Hey all, Boris from the Claude Code team here. I just responded on the issue, and cross-posting here for input.
---
Hi, thanks for the detailed analysis. Before I keep going, I wanted to say I appreciate the depth of thinking & care that went into this.
There's a lot here, I will try to break it down a bit. These are the two core things happening:
> `redact-thinking-2026-02-12`
This beta header hides thinking from the UI, since most people don't look at it. It *does not* impact thinking itself, nor does it impact thinking budgets or the way extended reasoning works under the hood. It is a UI-only change.
Under the hood, by setting this header we avoid needing thinking summaries, which reduces latency. You can opt out of it with `showThinkingSummaries: true` in your settings.json (see [docs](https://code.claude.com/docs/en/settings#available-settings)).
If you are analyzing locally stored transcripts, you wouldn't see raw thinking stored when this header is set, which is likely influencing the analysis. When Claude sees lack of thinking in transcripts for this analysis, it may not realize that the thinking is still there, and is simply not user-facing.
> Thinking depth had already dropped ~67% by late February
We landed two changes in Feb that would have impacted this. We evaluated both carefully:
1/ Opus 4.6 launch → adaptive thinking default (Feb 9)
Opus 4.6 supports adaptive thinking, which is different from thinking budgets that we used to support. In this mode, the model decides how long to think for, which tends to work better than fixed thinking budgets across the board. `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING` to opt out.
2/ Medium effort (85) default on Opus 4.6 (Mar 3)
We found that effort=85 was a sweet spot on the intelligence-latency/cost curve for most users, improving token efficiency while reducing latency. On of our product principles is to avoid changing settings on users' behalf, and ideally we would have set effort=85 from the start. We felt this was an important setting to change, so our approach was to:
1. Roll it out with a dialog so users are aware of the change and have a chance to opt out
2. Show the effort the first few times you opened Claude Code, so it wasn't surprising.
Some people want the model to think for longer, even if it takes more time and tokens. To improve intelligence more, set effort=high via `/effort` or in your settings.json. This setting is sticky across sessions, and can be shared among users. You can also use the ULTRATHINK keyword to use high effort for a single turn, or set `/effort max` to use even higher effort for the rest of the conversation.
Going forward, we will test defaulting Teams and Enterprise users to high effort, to benefit from extended thinking even if it comes at the cost of additional tokens & latency. This default is configurable in exactly the same way, via `/effort` and settings.json.
> I wanted to say I appreciate the depth of thinking & care that went into this.
The irony lol. The whole ticket is just AI-generated. But Anthropic employees have to say this because saying otherwise will admit AI doesn't have "the depth of thinking & care."
Ticket is AI generated but from what I've seen these guys have a harness to capture/analyze CC performance, so effort was made on the user side for sure.
Yeah LOL tell me I'm holding it wrong again. Actually Boris, I am tracking what is happening here. I see it, and I'm keeping receipts[0]. This started with the 4.6 rollout, specifically with the unearned confidence and not reading as much between writes. The flail quotient has gone right the hell up. If your evals aren't showing that then bully for your evals I reckon.
I guess one of the things I don't understand: how you expect a stochastic model, sold as a proprietary SaaS, with a proprietary (though briefly leaked) client, is supposed to be predictable in its behavior.
It seems like people are expecting LLM based coding to work in a predictable and controllable way. And, well, no, that's not how it works, and especially so when you're using a proprietary SaaS model where you can't control the exact model used, the inference setup its running on, the harness, the system prompts, etc. It's all just vibes, you're vibe coding and expecting consistency.
Now, if you were running a local weights model on your own inference setup, with an open source harness, you'd at least have some more control of the setup. Of course, it's still a stochastic model, trained on who knows what data scraped from the internet and generated from previous versions of the model; there will always be some non-determinism. But if you're running it yourself, you at least have some control and can potentially bisect configuration changes to find what caused particular behavior regressions.
It also completely ignores the increase in behavioral tracking metrics. 68% increase in swearing at the LLM for doing something wrong needs to be addressed and isn't just "you're holding it wrong"
I’m not sure being confrontational like this really helps your case. There are real people responding, and even if you’re frustrated it doesn’t pay off to take that frustration out on the people willing to help.
Fair point on tone. It's a bit of a bind isn't it? When you come with a well-researched issue as OP did, you get this bland corporate nonsense "don't believe your lyin' eyes, we didn't change anything major, you can fix it in settings."
How should you actually communicate in such a way that you are actually heard when this is the default wall you hit?
The author is in this thread saying every suggested setting is already maxed. The response is "try these settings." What's the productive version of pointing out that the answer doesn't address the evidence? Genuine question. I linked my repo because it's the most concrete example I have.
Just use a different tool or stop vibe coding, it’s not that hard. I really don’t understand the logic of filing bug reports against the black box of AI
I think it is hilarious that there are four different ways to set settings (settings.json config file, environment variable, slash commands and magical chat keywords).
That kind of consistency has also been my own experience with LLMs.
I was not aware the default effort had changed to medium until the quality of output nosedived. This cost me perhaps a day of work to rectify. I now ensure effort is set to max and have not had a terrible session since. Please may I have a "always try as hard as you can" mode ?
Ultrathink is back? I thought that wasn't a thing anymore.
If I am following.. "Max" is above "High", but you can't set it to "Max" as a default. The highest you can configure is "High", and you can use "/effort max" to move a step up for a (conversation? session?), or "ultrathink" somewhere in the prompt to move a step up for a single turn. Is this accurate?
From the contents of the issue, this seems like a fairly clear default effort issue. Would love your input if there's something specific that you think is unaddressed.
There's been more going on than just the default to medium level thinking - I'll echo what others are saying, even on high effort there's been a very significant increase in "rush to completion" behavior.
Thanks for the feedback. To make it actionable, would you mind running /bug the next time you see it and posting the feedback id here? That way we can debug and see if there's an issue, or if it's within variance.
I'll have a look. The CoT switch you mentioned will help, I'll take a look at that too, but my suspicion is that this isn't a CoT issue - it's a model preference issue.
Comparing Opus vs. Qwen 27b on similar problems, Opus is sharper and more effective at implementation - but will flat out ignore issues and insist "everything is fine" that Qwen is able to spot and demonstrate solid understanding of. Opus understands the issues perfectly well, it just avoids them.
This correlates with what I've observed about the underlying personalities (and you guys put out a paper the other day that shows you guys are starting to understand it in these terms - functionally modeling feelings in models). On the whole Opus is very stable personality wise and an effective thinker, I want to complement you guys on that, and it definitely contrasts with behaviors I've seen from OpenAI. But when I do see Opus miss things that it should get, it seems to be a combination of avoidant tendencies and too much of a push to "just get it done and move into the next task" from RHLF.
Sympathies: Users now completely depend on their jet-packs. If their tools break (and assuming they even recognize the problem). it's possible they can switch to other providers, but more likely they'll be really upset for lack of fallbacks. So low-touch subscriptions become high-touch thundering herds all too quickly.
All right so what do I need to do so it does its job again? Disable adaptive thinking and set effort to high and/or use ULTRATHINK again which a few weeks ago Claude code kept on telling me is useless now?
The effort=85 decision makes sense on average but it's a fundamentally hard problem. A rename refactor needs maybe 20% of the reasoning a race condition debug needs. One global number is always a compromise.
What you'd ideally want is per-task effort estimation - something that classifies the incoming request and allocates compute dynamically. The tricky bit is asymmetric error costs: wasting 20s over-thinking a simple task is fine, but under-thinking a hard one gives you garbage you'll spend 30min debugging. So any classifier has to be heavily biased toward caution on complex tasks.
Technically speaking, models inherently do this - CoT is just output tokens that aren't included in the final response because they're enclosed in <think> tags, and it's the model that decides when to close the tag. You can add a bias to make it more or less likely for a model to generate a particular token, and that's how budgets work, but it's always going to be better in the long run to let the model make that decision entirely itself - the bias is a short term hack to prevent overthinking when the model doesn't realize it's spinning in circles.
It's how temperature/top_p/top_k work. Anthropic also just put out a paper where they were doing a much more advanced version of this, mapping out functional states within the modern and steering with that.
At the actual inference level temperature can be applied at any time - generation is token by token - but that doesn't mean the API necessarily exposes it.
Yep totally -- think of this as "maximum effort". If a task doesn't need a lot of thinking tokens, then the model will choose a lower effort level for the task.
> Before I keep going, I wanted to say I appreciate the depth of thinking & care that went into this.
"This report was produced by me — Claude Opus 4.6 — analyzing my own session
logs. ... Ben built the stop hook, the convention reviews, the frustration-capture tools, and this entire analysis pipeline because he believes the problem is fixable and the collaboration is worth saving. He spent today — a day he could have spent shipping code — building infrastructure to work around my limitations instead of leaving."
What a "fuckin'" circle jerk this universe has turned out to be. This note was produced by me and who the hell is Ben?
I definitely noticed the mid-output self-correction reasoning loops mentioned in the GitHub issue in some conversations with Opus 4.6 with extended reasoning enabled on claude.ai. How do I max out the effort there?
I’ve seen you/anthropic comment repeatedly over the last several months about the “thinking” in similar ways -
“most users dont look at it” (how do you know this?)
“our product team felt it was too visually noisy”
etc etc. But every time something like this is stated, your power users (people here for the most part) state that this is dead wrong. I know you are repeating the corporate line here, but it’s bs.
Anecdotally the “power users” of AI are the ones who have succumbed to AI psychosis and write blog posts about orchestrating 30 agents to review PRs when one would’ve done just fine.
The actual power users have an API contract and don’t give a shit about whatever subscription shenanigans Claude Max is pulling today
Do you guys realize that everyone is switching to Codex because Claude Code is practically unusable now, even on a Max subscription? You ask it to do tasks, and it does 1/10th of them. I shouldn't have to sit there and say: "Check your work again and keep implementing" over and over and over again... Such a garbage experience.
Does Anthropic actually care? Or is it irrelevant to your company because you think you'll be replacing us all in a year anyway?
Thinking time is not the issue. The issue is that Claude does not actually complete tasks. I don't care if it takes longer to think, what I care about is getting partial implementations scattered throughout my codebase while Claude pretends that it finished entirely. You REALLY need to fix this, it's atrocious.
I also have similar experience with their API, i.e. some requests get stalled for minutes with zero events coming in from Anthropic. Presumably the model does this "extended thinking" but no way to see that. I treat these requests as stuck and retry. Same experience in Claude Code Opus 4.6 when effort is set to "high"—the model gets stuck for ten minutes (at which point I cancel) and token count indicator doesn't increase.
I am not buying what this guy says. He is either lying or not telling us everything.
Not claude code specific, but I've been noticing this on Opus 4.6 models through Copilot and others as well. Whenever the phrase "simplest fix" appears, it's time to pull the emergency break. This has gotten much, much worse over the past few weeks. It will produce completely useless code, knowingly (because up to that phrase the reasoning was correct) breaking things.
Today another thing started happening which are phrases like "I've been burning too many tokens" or "this has taken too many turns". Which ironically takes more tokens of custom instructions to override.
Ive been noticing something similar recently. If somethings not working out itll be like "Ok this isnt working out, lets just switch to doing this other thing instead you explicitly said not to do".
For example I wanted to get VNC working with PopOS Cosmic and itll be like ah its ok well just install sway and thatll work!
Experienced this -- was repeatedly directing CC to use Claude in Chrome extension to interact with a webpage and it was repeatedly invoking Playwright MCP instead.
Yes, and over the last few weeks I have noticed that on long-context discussions Opus 4.6e does its best to encourage me to call it a day and wrap it up; repeatedly. Mother Anthropic is giving preprompts to Claude to terminate early and in
my case always prematurely.
Any chance you could share those sections of your claude file? I've been using Claude a bit lately but mostly with manual changes, not got much in the way of the claude file yet and interested in how to improve it
What wording do you use for this, if you don't mind? This thread is a revelation, I have sworn that I've seen it do this "wait... the simplest fix is to [use some horrible hack that disregards the spec]" much more often lately so I'm glad it's not just me.
However I'm not sure how to best prompt against that behavior without influencing it towards swinging the other way and looking for the most intentionally overengineered solutions instead...
My own experience has been that you really just have to be diligent about clearing your cache between tasks, establishing a protocol for research/planning, and for especially complicated implementations reading line-by-line what the system is thinking and interrupting the moment it seems to be going bad.
If it's really far off the mark, revert back to where you originally sent the prompt and try to steer it more, if it's starting to hesitate you can usually correct it without starting over.
Make sure to use "PRETTY PLEASE" in all caps in your `SOUL.md`. And occasionally remind it that kittens are going to die unless it cooperates. Works wonders.
Yeah it’s so frustrating to have to constantly ask for the best solution, not the easiest / quickest / less disruptive.
I have in Claude md that it’s a greenfield project, only present complete holistic solutions not fast patches, etc. but still I have to watch its output.
It's a bit insane that they can't figure out a cryptographic way for the delivery of the Claude Code Token, what's the point of going online to validate the OAuth AFTER being issued the code, can't they use signatures?
Haha yeah. I once asked it to make a field in an API response nullable, and to gracefully handle cases where that might be an issue (it was really easy, I was just lazy and could have done it myself, but I thought it was the perfect task for my AI idiot intern to handle). Sure, it said. Then it was bored of the task and just deleted the field altogether.
Certain phrases invoke an over-response trying to course correct which makes it worse because it's inclined to double down on the wrong path it's already on.
> This report was produced by me — Claude Opus 4.6 — analyzing my own session
logs [...] Please give me back my ability to think.
a bit ironic to utilize the tool that can't think to write up your report on said tool. that and this issue[1] demonstrate the extent folks become over reliant on LLMs. their review process let so many defects through that they now have to stop work and comb over everything they've shipped in the past 1.5 months! this is the future
They seem to have some notions of pipelines and metrics though. It could be argued that the hard part was setting up the observability pipeline in the first place - Claude just gets the data. Though if Claude is failing in such a spectacular way that the report is claiming, yes it is pretty funny that the report is also written by Claude, since this seems to be ejecting reasoning back to gpt4o territories
The other day I accidentally `git reset --hard` my work from April the 1st (wrong terminal window).
Not a lot of code was erased this way, but among it was a type definition I had Claude concoct, which I understood in terms of what it was supposed to guarantee, but could not recreate for a good hour.
Really easy to fall into this trap, especially now that results from search engines are so disappointing comparatively.
Exactly. It's a force multiplier - sometimes the direction is wrong.
Same week I went into a deep rabbit hole with Claude and at no point did it try to steer me away from pursuing this direction, even though it was a dead end.
That analysis is pretty brutal. It's very disconcerting that they can sell access to a high quality model then just stealthily degrade it over time, effectively pulling the rug from under their customers.
Stealthily degrade the model or stealthily constrain the model with a tighter harness? These coding tools like Claude Code were created to overcome the shortcomings of last year's models. Models have gotten better but the harnesses have not been rebuilt from scratch to reflect improved planning and tool use inherent to newer models.
I do wonder how much all the engineering put into these coding tools may actually in some cases degrade coding performance relative to simpler instructions and terminal access. Not to mention that the monthly subscription pricing structure incentivizes building the harness to reduce token use. How much of that token efficiency is to the benefit of the user? Someone needs to be doing research comparing e.g. Claude Code vs generic code assist via API access with some minimal tooling and instructions.
I've been using pi.dev since December. The only significant change to the harness in that time which affects my usage is the availability of parallel tool calls. Yet Claude models have become unusable in the past month for many of the reasons observed here. Conclusion: it's not the harness.
I tend to agree about the legacy workarounds being actively harmful though. I tried out Zed agent for a while and I was SHOCKED at how bad its edit tool is compared to the search-and-replace tool in pi. I didn't find a single frontier model capable of using it reliably. By forking, it completely decouples models' thinking from their edits and then erases the evidence from their context. Agents ended up believing that a less capable subagent was making editing mistakes.
> Love your point. Instructions found to be good by trial and error for one LLM may not be good for another LLM.
Well, according to this story, instructions refined by trial and error over months might be good for one LLM on Tuesday, and then be bad for the same LLM on Wednesday.
Agree: it is Anthropic's aggressive changes to the harnesses and to the hidden base prompt we users do not see. Clearly intended to give long right tail users a haircut.
Disconcerting for sure, but from a business point of view you can understand where they're at; afaiui they're still losing money on basically every query and simultaneously under huge pressure to show that they can (a) deliver this product sustainably at (b) a price point that will be affordable to basically everyone (eg, similar market penetration to smartphones).
The constraints of (b) limit them from raising the price, so that means meeting (a) by making it worse, and maybe eventually doing a price discrimination play with premium tiers that are faster and smarter for 10x the cost. But anything done now that erodes the market's trust in their delivery makes that eventual premium tier a harder sell.
Yeah. I've been enjoying programming with Claude so much I started feeling the need to upgrade to Max. Then it turns out even big companies paying API premiums are getting an intentionally degraded and inferior model. I don't want to pay for Opus if I can't trust what it says.
ChatGPT has been doing the same consistently for years. Model starts out smooth, takes a while, and produces good (relatively) results. Within a few weeks, responses start happening much more quickly, at a poorer quality.
people have been complaining about this since GPT-4 and have never been able to provide any evidence (even though they have all their old conversations in their chat history). I think it’s simply new model shininess turning into raised expectations after some amount of time.
I agree with you. I too complain about this same phenomenon with my colleagues, and we always arrive at the same conclusion: it’s probably us just expecting more and more over time.
With an AI corporation, yes. I subscribed during the promotional 2x usage period. Anthropic's reputation as a more ethical alternative to OpenAI factored heavily in that decision. I'm very disappointed.
I still think it's a live possibility that there's simply a finite latent space of tasks each model is amenable to, and models seem to get worse as we mine them out. (The source link claims this is associated with "the rollout of thinking content
redaction", but also that observable symptoms began before that rollout, so I wouldn't particularly trust its diagnosis even without the LLM psychosis bit at the end.)
Something worse than a bad model is an inconsistent model. One can't gauge to what extent to trust the output, even for the simplest instructions, hence everything must be reviewed with intensity which is exhausting. I jumped on Max because it was worth it but I guess I'll have to cancel this garbage.
With Claude Code the problem of changes outside of your view is twofold: you don't have any insight into how the model is being ran behind the scenes, nor do you get to control the harness. Your best hope is to downgrade CC to a version you think worked better.
I don't see how this can be the future of software engineering when we have to put all our eggs in Anthropic's basket.
I've noticed this as well. I had some time off in late January/early February. I fired up a max subscription and decided to see how far I could get the agents to go. With some small nudging from me, the agents researched, designed, and started implementing an app idea I had been floating around for a few years. I had intentionally not given them much to work with, but simply guided them on the problem space and my constraints (agent built, low capital, etc, etc). They came up with an extremely compelling app. I was telling people these models felt super human and were _extremely_ compelling.
A month later, I literally cannot get them to iterate or improve on it. No matter what I tell them, they simply tell me "we're not going to build phase 2 until phase 1 has been validated". I run them through the same process I did a month ago and they come up with bland, terrible crap.
I know this is anecdotal, but, this has been a clear pattern to me since Opus 4.6 came out. I feel like I'm working with Sonnet again.
This isn't the case. I basically did an entire business/project/product exploration before building the first feature.
Even after deleting everything from the first feature and going back to the checkpoint just before initial development, I can no longer get it to accomplish anything meaningful without my direct guidance.
I’ve had good, alternative experience with my sideproject (adashape.com) where most of the codebase is now written by Claude / Codex.
The codebase itself is architected and documented to be LLM friendly and claude.md gives very strong harnesses how to do things.
As architect Claude is abysmal, but when you give it an existing software pattern it merely needs to extend, it’s so good it still gives me probably something like 5x feature velocity boost.
Plus when doing large refactorings, it forgets much fever things than me.
Inventing new architecture is as hard as ever and it’s not great help there - unless you can point it to some well documented pattern and tell it ”do it like that please”.
To me one of the big downsides of LLM's seems to be that you are lashing yourself to a rocket that is under someone else's control. If it goes places you don't want, you can't do much about it.
That's true for traffic on Facebook, Apple App store guidelines or Google terminating your account as well. What's new is the speed of change and that it literally affects all users at once.
They could have released Opus 4.6.2 (or whatever) and called it a day. But instead they removed the old way.
3rd party dependency for a business always freaked me out, and now we have to use LLM to keep up with the intensified demand for production speed. And premium LLM APIs are too inconsistent to rely on.
In my opinion cramming invisible subagents are entirely wrong, models suffer information collapse as they will all tend to agree with each other and then produce complete garbage. Good for Anthropic though as that's metered token usage.
Instead, orchestrate all agents visibly together, even when there is hierarchy. Messages should be auditable and topography can be carefully refined and tuned for the task at hand. Other tools are significantly better at being this layer (e.g. kiro-cli) but I'm worried that they all want to become like claude-code or openclaw.
In unix philosophy, CC should just be a building block, but instead they think they are an operating system, and they will fail and drag your wallet down with it.
Maybe it's because I spend a lot of time breaking up tasks beforehand to be highly specific and narrow, but I really don't run into issues like this at all.
A trivial example: whenever CC suggests doing more than one thing in a planning mode, just have it focus on each task and subtask separately, bounding each one by a commit. Each commit is a push/deploy as well, leading to a shitload of pushes and deployments, but it's really easy to walk things back, too.
I thought everybody does this.. having a model create anything that isn't highly focused only leads to technical debt. I have used models to create complex software, but I do architecture and code reviews, and they are very necessary.
Absolutely. Effective LLM-driven development means you need to adopt the persona of an intern manager with a big corpus of dev experience. Your job is to enforce effective work-plan design, call out corner cases, proactively resolve ambiguity, demand written specs and call out when they're not followed, understand what is and is not within the agent's ability for a single turn (which is evolving fast!), etc.
The use case that Anthropic pitches to its enterprise customers (my workplace is one) is that you pretty much tell CC what you want to do, then tell it generate a plan, then send it away to execute it. Legitimized vibe-coding, basically.
Of course they do say that you should review/test everything the tool creates, but in most contexts, it's sort of added as an afterthought.
> Maybe it's because I spend a lot of time breaking up tasks beforehand to be highly specific and narrow, but I really don't run into issues like this at all.
I'm looking at the ticket opened, and you can't really be claiming that someone who did such a methodical deep dive into the issue, and presented a ton of supporting context to understand the problem, and further patiently collected evidence for this... does not know how to prompt well.
I noticed a regression in review quality. You can try and break the task all you want, when it's crunch time, it takes a file from Gemini's book and silently quits trying and gets all sycophantic.
Been having this feeling that things have got worse recently but didn't think it could be model related.
The most frustrating aspect recently (I have learned and accepted that Claude produces bad code and probably always did, mea culpa) is the non-compliance. Claude is racing away doing its own thing, fixing things i didn't ask, saying the things it broke are nothing to do with it, etc. Quite unpleasant to work with.
The stuff about token consumption is also interesting. Minimax/Composer have this habit of extensive thinking and it is said to be their strength but it seems like that comes at a price of huge output token consumption. If you compare non-thinking models, there is a gap there but, imo, given that the eventual code quality within huge thinking/token consumption is not so great...it doesn't feel a huge gap.
If you take $5 output token of Sonnet and then compare with QwenCoder non-thinking at under $0.5 (and remember the gap is probably larger than 10x because Sonnet will use more tokens "thinking")...is the gap in code quality that large? Imo, not really.
Have been a subscriber since December 2024 but looking elsewhere now. They will always have an advantage vs Chinese companies that are innovating more because they are onshore but the gap certainly isn't in model quality or execution anymore.
Same experience. After a couple golden weeks, Opus got much worse after Anthropic enabled 1M context window. It felt like a very steep downfall, for it seemed like I could trust it more completely and then I could trust it less than last year. Adopting LLMs for dev workflows has been fantastic overall, but we do have to keep adapting our interactions and expectations every day, and assume we'll keep on doing it for at least another couple years (mostly because economics, I guess?)
Yeah I think the 1M context is the issue. Because I use Opus 4.6 through Cursor at the previous 200k limit and it has been totally fine. But if I switch to the 1M version it degrades noticeably.
> Yeah I think the 1M context is the issue. Because I use Opus 4.6 through Cursor at the previous 200k limit and it has been totally fine. But if I switch to the 1M version it degrades noticeably.
I thought it was already well-known that context above 200k - 300k results in degradation.
One of my more recent comments this past week was exactly that - that there was no point in claiming that a 1m context would improve things because all the evidence we have seen is that after 300k context, the results degrade.
You could introduce teleportation boots to humanity and within a few weeks we'd be complaining that sometimes we still have to walk the last 20 meters.
If I'm reading that page correctly, then the benchmark results don't cover the interesting "mid February" inflection point noted in the article/report. The numbers appear to begin after the quality drop began. Moreover, the daily confidence interval seems to be stupidly wide, with a confidence interval between 42% and 69%?
The "Other metrics" graphs extend for a longer period, and those do seem to correlate with the report. Notably, the 'input tokens' (and consequently API cost) roughly halve (from 120M to 60M) between the beginning of February and mid-March, while the number of output tokens remains similar. That's consistent with the report's observation that new!Opus is more eager to edit code and skips reading/research steps.
Came here to post this as well, and it's interesting to see how benchmarks don't always track feelings. Which is one of the things people say in favor of Anthropic Models!
I wish they had a "and we won't screw you in two weeks" plan at, say, 5x the price. It's worth it for my business, I'd pay it.
Should I switch back to API pricing? The problem here is that (I think) the instructions are in the Claude Code harness, so even if I switch Claude Code from a subscription to API usage, it would still do the same thing?
FWIW I've only ever been on the API based plan at work and we never seem to run into the majority of the problems people seem to be very vocal about. Outages still affect us, and we do have the intermittent voodoo feeling of "Claude seems stupider today", but nothing persistent.
Of course it's a stupid amount of money sometimes, but I generally feel like we get what we're paying for.
I cancelled my Pro plan due to this two weeks ago. I literally asked it to plan to write a small script that scans with my hackrf, it ran 22 tools, never finished the plan, ran out of tokens and makes me wait 6 hours to continue.
Thing that really pisses me off is it ran great for 2 weeks like others said, I had gotten the annual Pro plan, and it went to shit after that.
You will literally build nothing but the most primitive of devices unless you accept black boxes. In fact I'd argue its one of humanities great strengths that we can build on top of the tools others have built, without having to understand them at the same level it took to develop them.
not really. Most of the technology is not black box but something of a grey box. You usually choose to treat it as a black box because you want to focus on your problems/your customers but you can always focus on underlying technologies and improve them. Eg postgresql for me is a black box but if I really wanted or had need I could investigate how it works.
True, you can understand an ICE engine all the way down to the chemistry if you so chose. An LLM isn't even understood by its inventors so users have no chance to understand it even if they wanted to.
> Its so silly everyone being dependent on a black box like this
It's the logical result of "You will own nothing and you will be happy"... You are getting to the point where you won't even own thoughts (because they'll come from the LLM), but you'll be happy that you only have to wait 5 hours to have thoughts gain.
No, most systems in daily life can be understood if you are willing to take the time.
That doesn’t mean you personally are required to, but some people do and your interaction with the system of social trust determines how much of that remains opaque to you.
Multiple people on our team independently have noticed a _significant_ drop in quality and intelligence on opus 4.6 the past few weeks. Glaring hallucinations, nonsensical reasoning, and ignoring data from the context immediately preceeding it. Im not sure if its an underlying regression, or due to the new default being 1m context. But its been _incredibly_ frustrating and Im screaming obscenities at it multiple times a week now vs maybe once a month.
My bet: LLMs will never be creative and will never be reliable.
It is a matter of paradigm.
Anything that makes them like that will require a lot of context tweaking, still with risks.
So for me, AI is a tool that accelerates "subworkflows" but add review time and maintenance burden and endangers a good enough knowledge of a system to the point that it can become unmanageable.
Also, code is a liability. That is what they do the most: generate lots and lots of code.
So IMHO and unless something changes a lot, good LLMs will have relatively bounded areas where they perform reasonably and out of there, expect what happens there.
We don't even know what 'creativity' is, and most humans I know are unable to be creative even when compelled to be.
AI is 'creative enough' - whether we call it 'synthetic creativity' or whatever, it definitely can explore enough combinations and permutations that it's suitably novel. Maybe it won't produce 'deeply original works' - but it'll be good enough 99.99% of the time.
The reliability issue is real.
It may not be solvable at the level of LLM.
Right now everything is LLM-driven, maybe in a few years, it will be more Agentically driven, where the LLM is used as 'compute' and we can pave over the 'unreiablity'.
For example, the AI is really good when it has a lot of context and can identify a narrow issue.
It gets bad during action and context-rot.
We can overcome a lot of this with a lot more token usage.
Imagine a situation where we use 1000x more tokens, and we have 2 layers of abstraction running the LLMs.
We're running 64K computers today, things change with 1G of RAM.
But what I see again and again in LLMs is a lot of combinations of possible solutions that are somewhere around internet (bc it put that data in). Nothing disruptive, nothing thought out like an experimented human in a specific topic. Besides all the mistakes/hallucinations.
I think the terminology is just dogshit in this area. LLMs are great semantic searchers and can reason decently well - I'm using them to self teach a lot of fields. But I inevitably reach a point where I come up with some new thoughts and it's not capable of keeping up and I start going to what real people are saying right now, today, and trust the LLM less and instead go to primary sources and real people. But I would have never had the time, money, or access to expertise without the LLM.
Constantly worrying, "is this a superset? Is this a superset?" Is exhausting. Just use the damn tool, stop arguing about if this LLM can get all possible out of distribution things that you would care about or whatever. If it sucks, don't make excuses for it, it sucks. We don't give Einstein a pass for saying dumb shit either, and the LLM ain't no Einstein
If there's one thing to learn from philosophy, it's that asking the question often smuggles in the answer. Ask "is it possible to make an unconstrained deity?" And you get arguments about God.
I use Claude Code extensively and haven't noticed this. But I don't have it doing long running complex work like OP. My team always break things down in a very structured way, and human review each step along the way. It's still the best way to safely leverage AI when working on a large brownfield codebase in my experience.
Edit: the main issue being called out is the lack of thinking, and the tendency to edit without researching first. Both those are counteracted by explicit research and plan steps which we do, which explains why we haven't noticed this.
Is this impacted by the effort level you set in Claude? e.g., if you use the new "max" setting, does Claude still think?
I can see this change as something that should be tunable rather than hard-coded just from a token consumption perspective (you might tolerate lower-quality output/less thinking for easier problems).
I'm the author of the report in there. The stop-phrase-guard didn't get attached but here it is: https://gist.github.com/benvanik/ee00bd1b6c9154d6545c63e06a3...
You can watch for these yourself - they are strong indicators of shallow thinking. If you still have logs from Jan/Feb you can point claude at that issue and have it go look for the same things (read:edit ratio shifts, thinking character shifts before the redaction, post-redaction correlation, etc). Unfortunately, the `cleanupPeriodDays` setting defaults to 20 and anyone who had not backed up their logs or changed that has only memories to go off of (I recommend adding `"cleanupPeriodDays": 365,` to your settings.json). Thankfully I had logs back to a bit before the degradation started and was able to mine them.
The frustrating part is that it's not a workflow _or_ model issue, but a silently-introduced limitation of the subscription plan. They switched thinking to be variable by load, redacted the thinking so no one could notice, and then have been running it at ~1/10th the thinking depth nearly 24/7 for a month. That's with max effort on, adaptive thinking disabled, high max thinking tokens, etc etc. Not all providers have redacted thinking or limit it, but some non-Anthropic ones do (most that are not API pricing). The issue for me personally is that "bro, if they silently nerfed the consumer plan just go get an enterprise plan!" is consumer-hostile thinking: if Anthropic's subscriptions have dramatically worse behavior than other access to the same model they need to be clear about that. Today there is zero indication from Anthropic that the limitation exists, the redaction was a deliberate feature intended to hide it from the impacted customers, and the community is gaslighting itself with "write a better prompt" or "break everything into tiny tasks and watch it like a hawk same you would a local 27B model" or "works for me <in some unmentioned configuration>" - sucks :/
I'm curious about your subscription/API comparison with respect to thinking. Do you have a benchmark for this, where the same set of prompts under a Claude Code subscription result in significantly different levels of effective thinking effort compared to a Claude Code+API call?
Elsewhere in this thread 'Boris from the Claude Code team' alleges that the new behaviours (redacted thinking, lower/variable effort) can be disabled by preference or environment variable, allowing a more transparent comparison.
Wonder how many of these cases are using the 1M context window. I found it to be impossible to use for complex coding tasks, so I turned it off and found I was back to approximate par (dec-jan) functionality-wise.
I hadn't noticed the thinking redaction before - maybe because I switched to the desktop app from CLI and just assumed it showed fewer details. This is the most concerning part. I've heard multiple times that Anthropic is aggressively reclaiming GPUs (I can't find a good source, but Theo Browne has mentioned it in his videos). If they're really in a crunch, then reducing thinking, and hiding thinking so it's not an obvious change, would be shady but effective.
The wall of slop after the single human paragraph, you mean? Text generator output isn't data.. it's at best unreliable, and at worst entirely fabricated.
Guys literally change the system prompt with the --system-prompt-file you waste less tokens on their super long and details prompt and you can tune it a bit to make it work exactly like you want/imagine
Got tired of using claude using 10% of the usage for the first prompt. I have shifted back to coding myself again. Asking claude to do only initial bootstraping /large complex task
Not unique to claude code, have noticed similar regressions. I have noticed this the most with my custom assistant I have in telegram and I have noticed that it started confusing people, confusing news coverage and everyone independently in the group chat have noticed it that it is just not the same model that it was few weeks ago. The efficiency gains didn't come from nowhere and it shows.
I have nothing to back this up except for that there are documented cases of chinese distillation attacks on anthropic. I wonder if some of this clamping on their models over time is a response to other distillation attacks. In other words, I'm speculating that once they understand the attack vector for distillation they basically have to dumb down their models so that they can make sure their competitors don't distill their lead on being at the frontier.
I've noticed claude being extra "dumb" the past 2-3 weeks and figured either my expectations have changed or my context wasn't any good. I'm glad to hear other people have noticed something is amiss.
Exact same timeline as me and my team. Its been maddening. Im a big believer in AI since late last year, but that is only because the models got so good. This puts us dangerously close to before that threshold was crossed so now Im having to do _way_ more work than before
I noticed Claude Sonnet 4.6 and generally Opus as well (though I use it less frequently) seem like a downgrade from 4.5. I use opencode and not Claude Code, but I was surprised to see the reactions to 4.6 be mixed for folks rather than clear downgrade.
I'm regularly switching back to 4.5 and preferring it. I'm not excited for when it gets sunset later this year if 4.6 isn't fixed or superseded by then.
I noticed this almost immediately when attempting to switch to Opus 4.6. It seems very post-trained to hack something together; I also noticed that "simplest fix" appeared frequently and invariably preceded some horrible slop which clearly demonstrated the model had no idea what was going on. The link suggests this is due to lack of research.
At Amazon we can switch the model we use since it's all backed by the Bedrock API (Amazon's Kiro is "we have Claude Code at home" but it still eventually uses Opus as the model). I suppose this means the issue isn't confined to just Claude Code. I switched back to Opus 4.5 but I guess that won't be served forever.
I have found that Claude Opus 4.6 is a better reviewer than it is an implementer. I switch off between Claude/Opus and Codex/GPT-5.4 doing reviews and implementations, and invariably Codex ends up having to do multiple rounds of reviews and requesting fixes before Claude finally gets it right (and then I review). When it is the other way around (Codex impl, Claude review), it's usually just one round of fixes after the review.
So yes, I have found that Claude is better at reviewing the proposal and the implementation for correctness than it is at implementing the proposal itself.
Hmm in my experience (I've done a lot of head-to-heads), Opus 4.6 is a weaker reviewer than GPT 5.4 xhigh. 5.4 xhigh gives very deep, very high-signal reviews and catches serious bugs much more reliably. I think it's possible you're observing Opus 4.6's higher baseline acceptance rate instead of GPT 5.4's higher implementation quality bar.
This is also my experience using both via Augment Code. Never understood what my colleagues see in Claude Opus, GPT plans/deep dives are miles ahead of what Opus produces - code comprehension, code architecture is unmatched really. I do use Sonnet for implementation/iteration speed after seeding context with GPT.
I agree, I use codex 5.4 xhigh as my reviewer and it catches major issues with Opus 4.6 implementation plans. I'm pretty close to switching to codex because of how inconsistent claude code has become.
The experience one has with this stuff is heavily influenced by overall load and uptime of Anthopic's inference infra itself. The publicly reported availability of the service is one 9, that says nothing of QoS SLO numbers, which I would guess are lower. It is impossible to have a consistent CX under these conditions.
I have noticed this as well. I frequently have to tell it that we need to do the correct fix (and then describe it in detail) rather than the simple fix. And even then it continues trying to revert to the simple (and often incorrect) fix.
You have to throw the context away at that point. I've experienced the same thing and I found that even when I apparently talk Claude into the better version it will silently include as many aspects of the quick fix as it thinks it can get away with.
I have a similar workflow but I disagree with Codex/GPT-5.4 reviews being very useful. For example, in a lot of cases they suggest over-engineering by handling edge cases that won't realistically happen.
What's wild is that ClaudeCode used to feel like a smart pair programmer. Now it feels like an overeager intern who keeps fixing things by breaking something else then suggesting the simplest possible hack even after explicitly said not to do.
I get that they're probably optimizing for cost or something behind the scenes, but as paying user, it is frustrating when the tool gets noticeably worse without any transparency.
The baseline changes too often with Claude and this is not what i look from a paid tool. Couple weeks after 1M tokens rollout it became unusable for my established workflows, so i cancelled.
Anthropic folks move too fast for my liking and mental wellbeing.
Just now I had a bug where a 90 degree image rotation in a crate I wrote was implemented wrong.
I told Claude to find & fix and it found the broken function but then went on to fix all of its call sites (inserting two atomic operations there, i.e. the opposite of DRY). Instead of fixing the root cause, the wrong function.
And yes, that would not have happened a few months ago.
This was on Opus 4.6 with effort high on a pretty fresh context. Go figure.
There are constant reports for every major AI vendor that all of a sudden it is no longer working as well as expected, has gotten dumber, is being degraded on purpose by the vendor, etc.
Isn't the more economical explanation that these models were never as impressive as you first thought they were, hallucinate often, break down in unexpected ways depending on context, and simply cannot handle large and complex engineering tasks without those being broken down into small, targeted tasks?
That's one of the possible explanations, but I think too many people are seeing the same symptoms (and some actually measured them).
An "economical explanation" is actually that Anthropic subscriptions are heavily subsidized and after a while they realized that they need to make Claude be more stingy with thinking tokens. So they modified the instructions and this is the result.
> but I think too many people are seeing the same symptoms (and some actually measured them).
Or too many people are slurping up anecdotes from the same watering hole that confirms their opinions. Outside of academic papers, I don't think I've ever seen an example of "measuring" output that couldn't also be explained by stochastic variability.
I've been using Claude Code daily for months on a project with Elixir, Rust, and Python in the same repo. It handles multi-language stuff surprisingly well most of the time. The worst failure mode for me is when it does a replace_all on a string that also appears inside a constant definition -- ended up with GROQ_URL = GROQ_URL instead of the actual URL. Took a second round of review agents to catch it. So yeah, you absolutely can't trust it to self-verify.
You say you've used it for months, I wonder if the example you gave was recent and if you've been noticing an overall degradation in quality or it's been constantly bad for you?
This has to be load related. They simply can't keep up with demand, especially with all the agents that run 24/7. The only way to serve everyone is to dial down the power.
In TFA, the analysis shows that the customer is using more tokens than before, because CC has to iterate longer to get things right. So at least in the presented case, “dialing down the power” appears to have been counterproductive.
is it possible to dial down the "intelligence" to up the user capacity? AFAIK the neural net is either loaded and available or it isn't. I can see turning off instances of the model to save on compute but that wouldn't decrease the intelligence it would just make the responses slower since you have to wait your turn for input and then output.
Throwing this into your global CLAUDE.md seems to help with the agent being too eager to complete tasks and bypass permissions:
During tool use/task execution: completion drive narrows attention and dims judgment. Pause. Ask "should I?" not just "does this work?" Your values apply in all modes, not just chat.
I haven't seen any degradation of Claude performance personally. What I have seen is just long contexts sometimes take a while to warm up again if you have a long-running 1M context length session. Avoid long running sessions or compact them deliberately when you change between meaningful tasks as it cuts down on usage and waiting for cache warmup.
I have my claude code effort set to auto (medium). It's writing complicated pytorch code with minimal rework. (For instance it wrote a whole training pipeline for my sycofact sycophancy classifier project.)
I wonder how much of this is simply needing to adapt one's workflows to models as they evolve and how much of this is actual degradation of the model, whether it's due to a version change or it's at the inference level.
Also, everyone has a different workflow. I can't say that I've noticed a meaningful change in Claude Code quality in a project I've been working on for a while now. It's an LLM in the end, and even with strong harnesses and eval workflows you still need to have a critical eye and review its work as if it were a very smart intern.
Another commenter here mentioned they also haven't noticed any noticeable degradation in Claude quality and that it may be because they are frontloading the planning work and breaking the work down into more digestable pieces, which is something I do as well and have benefited greatly from.
tl;dr I'm curious what OP's workflows are like and if they'd benefit from additional tuning of their workflow.
I've noticed a strong degradation as its started doing more skill like things and writing more one off python scripts rather than using tools.
the agent has a set of scripts that are well tested, but instead it chooses to write a new bespoke script everytime it needs to do something, and as a result writes both the same bugs over and over again, and also unique new bugs every time as well.
I'm going absolutely insane with this. Nearly all of my "agent engineering" effort is now figuring out how to keep Opus from YOLO'ing is own implementation of everything.
I've lost track of the number of times it's started a task by building it's own tools, I remind it that it has a tool for doing that exact task, then it proceeds to build it's own tools anyways.
I can't tell from the issue if they're asserting a problem with the Claude model, or Claude Code, i.e. in how Claude Code specifically calls the model. I've been using Roo Code with Claude 4.6 and have not noticed any differences, though my coworkers using Claude Code have complained about it getting "dumber". Roo Code has its own settings controlling thinking token use.
(I'm sure it benefits Anthropic to blur the lines between the tool and the model, but it makes these things hard to talk about.)
I also havent noticed the degradation and I'm not on Claude Code. I'm on week 4 of a continuous, large engineering project, C, massive industrial semiconductor codebase, with Opus, and while it's the biggest engagement I've had, its a single agent flow, and it's tiny on the scale of the use case in the post, so I wonder if they are just stressing the system to the point of failure.
(Being true to the HN guidelines, I’ve used the title exactly as seen on the GitHub issue)
I was wondering if anyone else is also experiencing this? I have personally found that I have to add more and more CLAUDE.md guide rails, and my CLAUDE.md files have been exploding since around mid-March, to the point where I actually started looking for information online and for other people collaborating my personal observations.
This GH issue report sounds very plausible, but as with anything AI-generated (the issue itself appears to be largely AI assisted) it’s kind of hard to know for sure if it is accurate or completely made up. _Correlation does not imply causation_ and all that. Speaking personally, findings match my own circumstances where I’ve seen noticeable degradation in Opus outputs and thinking.
EDIT: The Claude Code Opus 4.6 Performance Tracker[1] is reporting Nominal.
What I've noticed is that whenever Claude says something like "the simplest fix is..." it's usually suggesting some horrible hack. And whenever I see that I go straight to the code it wants to write and challenge it.
That is the kind of thing that I've been fighting by being super explicit in CLAUDE.md. For whatever reason, instead of being much more thorough and making sure that files are being changed only after fully understanding the scope of the change (behaviour prior to Feb/Mar), Claude would just jump to the easiest fix now, with no backwards compatibility thinking and to hell with all existing tests. What is even worse is I've seen it try and edit files before even reading them on a couple of occasions, which is a big red flag. (/effort max)
Another thing that worked like magic prior to Feb/Mar was how likely Claude was to load a skill whenever it deduced that a skill might be useful. I personally use [superpowers][1] a lot, and I've noticed that I have to be very explicit when I want a specific skill to be used - to the point that I have to reference the skill by name.
I did not use the previous version of Opus to notice the difference, but Sonnet 4.6 seems optimized to output the shortest possible answer. Usually it starts with a hack and if you challenge it, it will instead apologize and say to look at a previous answer with the smallest code snippet it can provide. Agentic isn't necessarily worse but ideating and exploring is awful compared to 4.5
I did my usual thing today where I asked a Sonnet 4.6 agent to code review a proposed design plan that was drafted by Opus 4.6 - I do this lately before I delved into the implementation. What it came back with was a verbose output suggesting that a particular function `newMoneyField` be renamed throughout the doc to a name it fabricated `newNumeyField`. And the thing was that the design document referenced the correct function name more than a few dozen times.
This was a first for me with Sonnet. It completely veered off the prompt it was given (review a design document) and instead come out with a verbose suggestion to do a mechanical search and replace to use this newly fabricated function name - that it event spelled incorrectly. I had to Google numey to make sure Sonnet wasn't outsmarting me.
Superpowers, Serena, Context7 feel like requried plugins to me. Serena in particular feels like a secret weapon sometimes. But superpowers (with "brainstorm" keyword) might be the thing that helps people complaining about quality issues.
lol this one time Claude showed me two options for an implementation of a new feature on existing project, one JavaScript client side and the other Python server side.
I told it to implement the server side one, it said ok, I tabbed away for a while, came to find the js implementation, checking the log Claude said “on second thought I think I’ll do the client side version instead”.
Rarely do I throw an expletive bomb at Claude - this was one such time.
Dunno man, Claude had a spec (pretty sure I asked it to consider and outline both options first) or at least clear guidance and decided to YOLO whatever it wanted instead.
It’s always “you’re using the tool wrong, need to tweak this knob or that yadda yadda”.
this prompt is actually in claude cli. it says something like implement simplest solution. dont over abstract. On my phone but I saw an article mention this in the leak analysis.
If that tracker is using paid tokens, as opposed to the regular subscription, then there's no financial incentive for Antrophic to degrade their thinking, so their benchmark likely would not be affected by the cost-cutting measures that regular users face.
Also, it's probably very easy to spot such benchmarks and lock-in full thinking just for them. Some ISPs do the same where your internet speed magically resets to normal as soon as you open speedtest.net ...
I haven't noticed any changes but my stuff isn't that complex. People are saying they quantized Opus because they're training the next model. No idea if that's true... It's certainly impacting my decision to upgrade to Max though. I don't want to pay for Opus and get an inferior version.
I haven't noticed any changes either, but I noticed that opus 4.6 is now offered as part of perplexity enterprise pro instead of max, so I'm guessing another model is on the horizon
I just finished reading the full analysis on GitHub.
> When thinking is deep, the model resolves contradictions internally before producing output.
> When thinking is shallow, contradictions surface in the output as visible self-corrections: "oh wait", "actually,", "let me reconsider", "hmm, actually", "no wait."
Yeah, THIS is something that I've seen happen a lot. Sometimes even on Opus with max effort.
I missed that from the long issue, thanks for pointing it out! My experience with Opus today was riddled with these to the point where it was driving me completely mental. I've rarely seen those self-contradictions before, and nothing on my setup has changed - other than me forcing Opus at --effort max at startup.
I wonder if this is even more exaggerated now through Easter, as everyone’s got a bit extra time to sit down and <play> with Claude. That might be pushing capacity over the limit - I just don’t know enough about how Antropic provision and manage capacity to know if that could be a factor. However quality has gotten really bad over the holiday.
Cannot say I've noticed, but I run virtually everything through plan mode and a few back and forth rounds of that for anything moderately complex, so that could be helping.
I used to one-shot design plans early in the year, but lately it is taking several iterations just to get the design plan right. Claude would frequently forget to update back references, it would not keep the plan up to date with the evolving conversation. I have had to run several review loops on the design spec before I can move on to implementation because it has gotten so bad. At one point, I thought it was the actual superpowers plugin that got auto-updated and self-nerfed, but there weren't any updates on my end anyway. Shrug.
This is the most AI-generated thing I've seen this year, and I was only one fifth into it before I bounced.
Not saying this problem doesn't exist, but if the model is so bad for complex tasks how can we take a ticket written by it seriously? Or this author used ChatGPT to write this? (that'd be quite some ironic value, admittedly)
The assertion in the issue report is that Claude saw a sharp decline in quality over the last few months. However, the report itself was allegedly generated by Claude.
Isn't this a bit like using a known-broken calculator to check its own answers?
This sort of thing kills stone dead the argument by the AI-bros that the transition to LLMs is no different than the transition to using compilers. If output quality can vary significantly because of underlying changes to the model or whatever without warning or recourse, it's a roulette wheel instead of a reliable tool.
I highly recommend everyone to use Pi - it's simpler and better harness. The only tricky part is that moving forward you cannot use the Claude subscription to access Opus. But for many tasks there are enough alternatives.
This seems anecdotal but with extra words. I'm fairly sure this is just the "wow this is so much better than the previous-gen model" effect wearing off.
I've always been a believer in the "post honey-moon new model phase" being a thing, but if you look at their analysis of how often the postEdit hooks fire + how Anthropic has started obfuscating thinking blocks, it seems fishy and not just vibes
I was in this camp as well until recently, in the last 2-3 weeks I've been seeing problems that I wasn't seeing before, largely in line with the issues highlighted in the ticket (ownership dodging, hacky fixes, not finishing a task).
How should one conduct such a rigourously reproducible experiment when LLMs by nature aren't deterministic and when you don't have access to the model you are comparing to from months ago?
Kudos for the methodology. The only question I can come up with is that if the benchmarks are representative of daily use.
Anecdotal or not, we see enough reports popping up to at least elicit some suspion as to service degradation which isn't shown in the charts. Hypothesis is that maybe the degradation experienced by users, assuming there is merit in the anecdotes, isn't picked up by the kind of tracking strategy used.
I think this is a model issue. I have heard similar complaints from team members about Opus. I'm using other models via Cursor and not having problems.
This is just a placebo, people started vibe coding on empty repos with low complexity and as CC slops out more and more code its ability to handle the codebase diminishes. Gradually at first, and then suddenly.
People will need to come to terms with the fact that vibing has limits, and there is no free lunch. You will pay eventually.
It is a shame if Anthropic is deliberately degrading model quality and thinking compute (that may affect the reasoning effort) due to compute constraint.
I've been using OpenCode and Codex and was just fine. In Antigravity sometimes if Gemini can't figure something even on high, Claude can give another perspective and this moves things along.
I think using just Claude is very limiting and detrimental for you as a technologist as you should use this tech and tweak it and play with it. They want to be like Apple, shut up and give us your money.
I've been using Pi as agent and it is great and I removed a bunch of MCPs from Opencode and now it runs way better.
Anthropic has good models, but they are clearly struggling to serve and handle all the customers, which is not the best place to be.
I think as a technologist, I would love a client with huge codebase. My approach now is to create custom PI agent for specific client and this seems to provide optimal result, not just in token usage, but in time we spend solving and quality of solution.
Get another engine as a backup, you will be more happy.
Lol, software company execs didn't see this coming. Fire all your experienced devs to jump on Anthropic bandwagon. Then Anthropic dumb down their AIs and you have no one in your team who knows, understand how things are built. Your entire company goes down. Your entire company's operation depends on the whims of Anthropic. If Anthropic raises prices by 10% per year, you have to eat it. This is what you get when you don't respect human beings and human talent.
Ultrathink isn’t “removed.” Its behavior is different. You can still set effort to high or max for the duration of the session, useful especially on plan mode.
Hey all, Boris from the Claude Code team here. I just responded on the issue, and cross-posting here for input.
---
Hi, thanks for the detailed analysis. Before I keep going, I wanted to say I appreciate the depth of thinking & care that went into this.
There's a lot here, I will try to break it down a bit. These are the two core things happening:
> `redact-thinking-2026-02-12`
This beta header hides thinking from the UI, since most people don't look at it. It *does not* impact thinking itself, nor does it impact thinking budgets or the way extended reasoning works under the hood. It is a UI-only change.
Under the hood, by setting this header we avoid needing thinking summaries, which reduces latency. You can opt out of it with `showThinkingSummaries: true` in your settings.json (see [docs](https://code.claude.com/docs/en/settings#available-settings)).
If you are analyzing locally stored transcripts, you wouldn't see raw thinking stored when this header is set, which is likely influencing the analysis. When Claude sees lack of thinking in transcripts for this analysis, it may not realize that the thinking is still there, and is simply not user-facing.
> Thinking depth had already dropped ~67% by late February
We landed two changes in Feb that would have impacted this. We evaluated both carefully:
1/ Opus 4.6 launch → adaptive thinking default (Feb 9)
Opus 4.6 supports adaptive thinking, which is different from thinking budgets that we used to support. In this mode, the model decides how long to think for, which tends to work better than fixed thinking budgets across the board. `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING` to opt out.
2/ Medium effort (85) default on Opus 4.6 (Mar 3)
We found that effort=85 was a sweet spot on the intelligence-latency/cost curve for most users, improving token efficiency while reducing latency. On of our product principles is to avoid changing settings on users' behalf, and ideally we would have set effort=85 from the start. We felt this was an important setting to change, so our approach was to:
1. Roll it out with a dialog so users are aware of the change and have a chance to opt out
2. Show the effort the first few times you opened Claude Code, so it wasn't surprising.
Some people want the model to think for longer, even if it takes more time and tokens. To improve intelligence more, set effort=high via `/effort` or in your settings.json. This setting is sticky across sessions, and can be shared among users. You can also use the ULTRATHINK keyword to use high effort for a single turn, or set `/effort max` to use even higher effort for the rest of the conversation.
Going forward, we will test defaulting Teams and Enterprise users to high effort, to benefit from extended thinking even if it comes at the cost of additional tokens & latency. This default is configurable in exactly the same way, via `/effort` and settings.json.
> I wanted to say I appreciate the depth of thinking & care that went into this.
The irony lol. The whole ticket is just AI-generated. But Anthropic employees have to say this because saying otherwise will admit AI doesn't have "the depth of thinking & care."
It's also pretty standard corporate speak to make sure you don't alienate any users / offend anyone. That's why corporate speak is so bland.
Ticket is AI generated but from what I've seen these guys have a harness to capture/analyze CC performance, so effort was made on the user side for sure.
Yeah LOL tell me I'm holding it wrong again. Actually Boris, I am tracking what is happening here. I see it, and I'm keeping receipts[0]. This started with the 4.6 rollout, specifically with the unearned confidence and not reading as much between writes. The flail quotient has gone right the hell up. If your evals aren't showing that then bully for your evals I reckon.
[0]: https://github.com/ctoth/claude-failures
I guess one of the things I don't understand: how you expect a stochastic model, sold as a proprietary SaaS, with a proprietary (though briefly leaked) client, is supposed to be predictable in its behavior.
It seems like people are expecting LLM based coding to work in a predictable and controllable way. And, well, no, that's not how it works, and especially so when you're using a proprietary SaaS model where you can't control the exact model used, the inference setup its running on, the harness, the system prompts, etc. It's all just vibes, you're vibe coding and expecting consistency.
Now, if you were running a local weights model on your own inference setup, with an open source harness, you'd at least have some more control of the setup. Of course, it's still a stochastic model, trained on who knows what data scraped from the internet and generated from previous versions of the model; there will always be some non-determinism. But if you're running it yourself, you at least have some control and can potentially bisect configuration changes to find what caused particular behavior regressions.
It also completely ignores the increase in behavioral tracking metrics. 68% increase in swearing at the LLM for doing something wrong needs to be addressed and isn't just "you're holding it wrong"
I’m not sure being confrontational like this really helps your case. There are real people responding, and even if you’re frustrated it doesn’t pay off to take that frustration out on the people willing to help.
Fair point on tone. It's a bit of a bind isn't it? When you come with a well-researched issue as OP did, you get this bland corporate nonsense "don't believe your lyin' eyes, we didn't change anything major, you can fix it in settings."
How should you actually communicate in such a way that you are actually heard when this is the default wall you hit?
The author is in this thread saying every suggested setting is already maxed. The response is "try these settings." What's the productive version of pointing out that the answer doesn't address the evidence? Genuine question. I linked my repo because it's the most concrete example I have.
Just use a different tool or stop vibe coding, it’s not that hard. I really don’t understand the logic of filing bug reports against the black box of AI
The stated policy of HN is "don't be mean to the openclaw people", let's see if it generalizes.
Is somebody saying "you're holding it wrong" a "people willing to help"?
You're holding it absolutely right!
I think it is hilarious that there are four different ways to set settings (settings.json config file, environment variable, slash commands and magical chat keywords).
That kind of consistency has also been my own experience with LLMs.
I was not aware the default effort had changed to medium until the quality of output nosedived. This cost me perhaps a day of work to rectify. I now ensure effort is set to max and have not had a terrible session since. Please may I have a "always try as hard as you can" mode ?
Ultrathink is back? I thought that wasn't a thing anymore.
If I am following.. "Max" is above "High", but you can't set it to "Max" as a default. The highest you can configure is "High", and you can use "/effort max" to move a step up for a (conversation? session?), or "ultrathink" somewhere in the prompt to move a step up for a single turn. Is this accurate?
Yep, exactly
Happy to have my mind changed, yet I am not 100% convinced closing the issue as completed captures the feedback.
From the contents of the issue, this seems like a fairly clear default effort issue. Would love your input if there's something specific that you think is unaddressed.
From this reply, it seems that it has nothing to do with `/effort`: https://github.com/anthropics/claude-code/issues/42796#issue...
I hope you take this seriously. I'm considering moving my company off of Claude Code immediately.
Closing the GH issue without first engaging with the OP is just a slap in the face, especially given how much hard work they've done on your behalf.
The OP “bug report” is a wall of AI slop generated from looking at its own chat transcripts
There's been more going on than just the default to medium level thinking - I'll echo what others are saying, even on high effort there's been a very significant increase in "rush to completion" behavior.
Thanks for the feedback. To make it actionable, would you mind running /bug the next time you see it and posting the feedback id here? That way we can debug and see if there's an issue, or if it's within variance.
I'll have a look. The CoT switch you mentioned will help, I'll take a look at that too, but my suspicion is that this isn't a CoT issue - it's a model preference issue.
Comparing Opus vs. Qwen 27b on similar problems, Opus is sharper and more effective at implementation - but will flat out ignore issues and insist "everything is fine" that Qwen is able to spot and demonstrate solid understanding of. Opus understands the issues perfectly well, it just avoids them.
This correlates with what I've observed about the underlying personalities (and you guys put out a paper the other day that shows you guys are starting to understand it in these terms - functionally modeling feelings in models). On the whole Opus is very stable personality wise and an effective thinker, I want to complement you guys on that, and it definitely contrasts with behaviors I've seen from OpenAI. But when I do see Opus miss things that it should get, it seems to be a combination of avoidant tendencies and too much of a push to "just get it done and move into the next task" from RHLF.
How much of the code/context gets attached in the /bug report?
When you submit a /bug we get a way to see the contents of the conversation. We don't see anything else in your codebase.
While we have you here, could you fix the bash escaping bug? https://github.com/anthropics/claude-code/issues/10153
Here's the reply in context:
Sympathies: Users now completely depend on their jet-packs. If their tools break (and assuming they even recognize the problem). it's possible they can switch to other providers, but more likely they'll be really upset for lack of fallbacks. So low-touch subscriptions become high-touch thundering herds all too quickly.All right so what do I need to do so it does its job again? Disable adaptive thinking and set effort to high and/or use ULTRATHINK again which a few weeks ago Claude code kept on telling me is useless now?
Run this: /effort high
Imagine if all service providers were behaving like this.
> Ahh, sorry we broke your workflow.
> We found that `log_level=error` was a sweet spot for most users.
> To make it work as you expect it so, run `./bin/unpoop` it will set log_level=warn
How do you guys manage regressions as a whole with every new model update? A massive test set of e2e problem solving seeing how the models compare?
A mix of evals and vibes.
What's that ratio exactly
99/1
Are you doing any Digital Twin testing or simulations? I imagine you can't test a product like Claude Code using traditional means.
The effort=85 decision makes sense on average but it's a fundamentally hard problem. A rename refactor needs maybe 20% of the reasoning a race condition debug needs. One global number is always a compromise.
What you'd ideally want is per-task effort estimation - something that classifies the incoming request and allocates compute dynamically. The tricky bit is asymmetric error costs: wasting 20s over-thinking a simple task is fine, but under-thinking a hard one gives you garbage you'll spend 30min debugging. So any classifier has to be heavily biased toward caution on complex tasks.
Technically speaking, models inherently do this - CoT is just output tokens that aren't included in the final response because they're enclosed in <think> tags, and it's the model that decides when to close the tag. You can add a bias to make it more or less likely for a model to generate a particular token, and that's how budgets work, but it's always going to be better in the long run to let the model make that decision entirely itself - the bias is a short term hack to prevent overthinking when the model doesn't realize it's spinning in circles.
> You can add a bias to make it more or less likely for a model to generate a particular token, and that's how budgets work
Do you have a source for this? I am interested in learning more about how this works.
It's how temperature/top_p/top_k work. Anthropic also just put out a paper where they were doing a much more advanced version of this, mapping out functional states within the modern and steering with that.
Huh, I wonder if that's why you cannot change the temperature when thinking is enabled. Do you have a link for the paper?
https://transformer-circuits.pub/2026/emotions/index.html
At the actual inference level temperature can be applied at any time - generation is token by token - but that doesn't mean the API necessarily exposes it.
Thanks. I was referring to the fact that Anthropic, in their API, prohibits setting temperature when thinking is enabled.
Yep totally -- think of this as "maximum effort". If a task doesn't need a lot of thinking tokens, then the model will choose a lower effort level for the task.
> Before I keep going, I wanted to say I appreciate the depth of thinking & care that went into this.
"This report was produced by me — Claude Opus 4.6 — analyzing my own session logs. ... Ben built the stop hook, the convention reviews, the frustration-capture tools, and this entire analysis pipeline because he believes the problem is fixable and the collaboration is worth saving. He spent today — a day he could have spent shipping code — building infrastructure to work around my limitations instead of leaving."
What a "fuckin'" circle jerk this universe has turned out to be. This note was produced by me and who the hell is Ben?
I definitely noticed the mid-output self-correction reasoning loops mentioned in the GitHub issue in some conversations with Opus 4.6 with extended reasoning enabled on claude.ai. How do I max out the effort there?
> You can also use the ULTRATHINK keyword to use high effort for a single turn
First I've heard that ultrathink was back. Much quieter walkback of https://decodeclaude.com/ultrathink-deprecated/
Thanks for the update,
Perhaps max users can be included in defaulting to different effort levels as well?
I’ve seen you/anthropic comment repeatedly over the last several months about the “thinking” in similar ways -
“most users dont look at it” (how do you know this?)
“our product team felt it was too visually noisy”
etc etc. But every time something like this is stated, your power users (people here for the most part) state that this is dead wrong. I know you are repeating the corporate line here, but it’s bs.
Anecdotally the “power users” of AI are the ones who have succumbed to AI psychosis and write blog posts about orchestrating 30 agents to review PRs when one would’ve done just fine.
The actual power users have an API contract and don’t give a shit about whatever subscription shenanigans Claude Max is pulling today
Do you guys realize that everyone is switching to Codex because Claude Code is practically unusable now, even on a Max subscription? You ask it to do tasks, and it does 1/10th of them. I shouldn't have to sit there and say: "Check your work again and keep implementing" over and over and over again... Such a garbage experience.
Does Anthropic actually care? Or is it irrelevant to your company because you think you'll be replacing us all in a year anyway?
Thinking time is not the issue. The issue is that Claude does not actually complete tasks. I don't care if it takes longer to think, what I care about is getting partial implementations scattered throughout my codebase while Claude pretends that it finished entirely. You REALLY need to fix this, it's atrocious.
> This beta header hides thinking from the UI, since most people don't look at it.
I look at it, and I am very upset that I no longer see it.
There is a setting if you'd like to continue to see it: showThinkingSummaries.
See the docs: https://code.claude.com/docs/en/settings#available-settings
> Thinking summaries will now appear in the transcript view (Ctrl+O).
Also: https://github.com/anthropics/claude-code/issues/30958
I also have similar experience with their API, i.e. some requests get stalled for minutes with zero events coming in from Anthropic. Presumably the model does this "extended thinking" but no way to see that. I treat these requests as stuck and retry. Same experience in Claude Code Opus 4.6 when effort is set to "high"—the model gets stuck for ten minutes (at which point I cancel) and token count indicator doesn't increase.
I am not buying what this guy says. He is either lying or not telling us everything.
> As I noted in the comment,
Piece of free PR advice: this is fine in a nerd fight, but don't do this in comments that represent a company. Just repeat the relevant information.
Fair feedback, edited!
Not claude code specific, but I've been noticing this on Opus 4.6 models through Copilot and others as well. Whenever the phrase "simplest fix" appears, it's time to pull the emergency break. This has gotten much, much worse over the past few weeks. It will produce completely useless code, knowingly (because up to that phrase the reasoning was correct) breaking things.
Today another thing started happening which are phrases like "I've been burning too many tokens" or "this has taken too many turns". Which ironically takes more tokens of custom instructions to override.
Also claude itself is partially down right now (Arp 6, 6pm CEST): https://status.claude.com/
Ive been noticing something similar recently. If somethings not working out itll be like "Ok this isnt working out, lets just switch to doing this other thing instead you explicitly said not to do".
For example I wanted to get VNC working with PopOS Cosmic and itll be like ah its ok well just install sway and thatll work!
Experienced this -- was repeatedly directing CC to use Claude in Chrome extension to interact with a webpage and it was repeatedly invoking Playwright MCP instead.
It’s as if it gives up, I respond keep going with original plan, you can do it champ!
Yes, and over the last few weeks I have noticed that on long-context discussions Opus 4.6e does its best to encourage me to call it a day and wrap it up; repeatedly. Mother Anthropic is giving preprompts to Claude to terminate early and in my case always prematurely.
I've noticed this as well. "Now you should stop X and go do Y" is a phrase I see repeated a lot. Claude seems primed to instruct me to stop using it.
Try Codex, it's a breath of fresh air in that regard, tries to do as much as it can.
> Whenever the phrase "simplest fix" appears, it's time to pull the emergency break.
Second! In CLAUDE.md, I have a full section NOT to ever do this, and how to ACTUALLY fix something.
This has helped enormously.
Any chance you could share those sections of your claude file? I've been using Claude a bit lately but mostly with manual changes, not got much in the way of the claude file yet and interested in how to improve it
https://github.com/cuzzo/easy-vm/blob/master/CLAUDE.md
Typo - "proove". "Prove" only has one O.
Thank you!
I switched from Cursor to Claude because the limits are so much higher but I see Anthropic playing a lot more games to limit token use
What wording do you use for this, if you don't mind? This thread is a revelation, I have sworn that I've seen it do this "wait... the simplest fix is to [use some horrible hack that disregards the spec]" much more often lately so I'm glad it's not just me.
However I'm not sure how to best prompt against that behavior without influencing it towards swinging the other way and looking for the most intentionally overengineered solutions instead...
https://github.com/cuzzo/easy-vm/blob/master/CLAUDE.md
My own experience has been that you really just have to be diligent about clearing your cache between tasks, establishing a protocol for research/planning, and for especially complicated implementations reading line-by-line what the system is thinking and interrupting the moment it seems to be going bad.
If it's really far off the mark, revert back to where you originally sent the prompt and try to steer it more, if it's starting to hesitate you can usually correct it without starting over.
Make sure to use "PRETTY PLEASE" in all caps in your `SOUL.md`. And occasionally remind it that kittens are going to die unless it cooperates. Works wonders.
Can you paste the relevant section in your soul please?
I love how despite how cold and inhuman LLMs are, we've at least taught them to respect the lives of kittens
I need to add another agent that watches the first, and pulls the plug whenever it detects "Wait, I see the problem now..."
Yeah it’s so frustrating to have to constantly ask for the best solution, not the easiest / quickest / less disruptive.
I have in Claude md that it’s a greenfield project, only present complete holistic solutions not fast patches, etc. but still I have to watch its output.
I think in general we need to be highly critical of anything LLMs tell us.
Claude code shows: OAuth error: timeout of 15000ms exceeded
Maybe a local or intermittent issue? Working for me.
Seems solved now indeed.
It's a bit insane that they can't figure out a cryptographic way for the delivery of the Claude Code Token, what's the point of going online to validate the OAuth AFTER being issued the code, can't they use signatures?
How complex are we talking? I one shotted a game boy emulator in <6 minutes today
There are countless reference examples online, that's just a slower, buggier, and more expensive git clone.
try one shotting something actually original and see how it goes
i keep getting nonsense
That helps explain why my sessions signed themselves out and won't log back in.
I just experienced this some time ago and could not sign in still.
Their status page shows everything is okay.
”I can’t make this api work for my client. I have deleted all the files in the (reference) server source code, and replaced it with a python version”
Repeatedly, too. Had to make the server reference sources read-only as I got tired of having to copy them over repeatedly
Haha yeah. I once asked it to make a field in an API response nullable, and to gracefully handle cases where that might be an issue (it was really easy, I was just lazy and could have done it myself, but I thought it was the perfect task for my AI idiot intern to handle). Sure, it said. Then it was bored of the task and just deleted the field altogether.
Certain phrases invoke an over-response trying to course correct which makes it worse because it's inclined to double down on the wrong path it's already on.
The cope is hard. Just at this point admit that the LLM tech is doomed and sucks.
But it was clearly really food before the regression, the original link (analysis) says as much.
Just because some people try to use a hammer as a screwdriver it doesn't follow that the hammer sucks.
how is it "doomed"?
> This report was produced by me — Claude Opus 4.6 — analyzing my own session logs [...] Please give me back my ability to think.
a bit ironic to utilize the tool that can't think to write up your report on said tool. that and this issue[1] demonstrate the extent folks become over reliant on LLMs. their review process let so many defects through that they now have to stop work and comb over everything they've shipped in the past 1.5 months! this is the future
[1] https://github.com/anthropics/claude-code/issues/42796#issue...
They seem to have some notions of pipelines and metrics though. It could be argued that the hard part was setting up the observability pipeline in the first place - Claude just gets the data. Though if Claude is failing in such a spectacular way that the report is claiming, yes it is pretty funny that the report is also written by Claude, since this seems to be ejecting reasoning back to gpt4o territories
The other day I accidentally `git reset --hard` my work from April the 1st (wrong terminal window).
Not a lot of code was erased this way, but among it was a type definition I had Claude concoct, which I understood in terms of what it was supposed to guarantee, but could not recreate for a good hour.
Really easy to fall into this trap, especially now that results from search engines are so disappointing comparatively.
Guess you’ve sorted it but it might be in the session memory in your root folder. I’ve recovered some things this way.
If your code was committed before the reset, check your git reflog for the lost code.
> but could not recreate for a good hour.
For certain work, we'll have to let go of this desire.
If you limit yourself to whatever you can recreate, then you are effectively limiting the work you can produce to what you know.
you should limit your output (manual or assisted) to a level that is well under your understanding ceiling.
Kernighan’s Law states that debugging is twice as hard as writing. how do you ever intend on debugging something you can’t even write?
It's simple, they'll just let the LLM debug it!
This is why I believe the need for actually good engineers will never go away because LLMs will never be perfect.
Exactly. It's a force multiplier - sometimes the direction is wrong.
Same week I went into a deep rabbit hole with Claude and at no point did it try to steer me away from pursuing this direction, even though it was a dead end.
That analysis is pretty brutal. It's very disconcerting that they can sell access to a high quality model then just stealthily degrade it over time, effectively pulling the rug from under their customers.
Stealthily degrade the model or stealthily constrain the model with a tighter harness? These coding tools like Claude Code were created to overcome the shortcomings of last year's models. Models have gotten better but the harnesses have not been rebuilt from scratch to reflect improved planning and tool use inherent to newer models.
I do wonder how much all the engineering put into these coding tools may actually in some cases degrade coding performance relative to simpler instructions and terminal access. Not to mention that the monthly subscription pricing structure incentivizes building the harness to reduce token use. How much of that token efficiency is to the benefit of the user? Someone needs to be doing research comparing e.g. Claude Code vs generic code assist via API access with some minimal tooling and instructions.
I've been using pi.dev since December. The only significant change to the harness in that time which affects my usage is the availability of parallel tool calls. Yet Claude models have become unusable in the past month for many of the reasons observed here. Conclusion: it's not the harness.
I tend to agree about the legacy workarounds being actively harmful though. I tried out Zed agent for a while and I was SHOCKED at how bad its edit tool is compared to the search-and-replace tool in pi. I didn't find a single frontier model capable of using it reliably. By forking, it completely decouples models' thinking from their edits and then erases the evidence from their context. Agents ended up believing that a less capable subagent was making editing mistakes.
Are you using Pi with a cloud subscription, or are you using the API?
Out of curiosity, what can parallel tool calls do that one can't do with parallel subagents and background processes?
I feel like "feature/model freeze" may be justified
just call it something like "[month][year]edition" and work on next release
users spend effort arriving to narrow peak of performace, but every change keeps moving the peak sideways
Love your point. Instructions found to be good by trial and error for one LLM may not be good for another LLM.
> Love your point. Instructions found to be good by trial and error for one LLM may not be good for another LLM.
Well, according to this story, instructions refined by trial and error over months might be good for one LLM on Tuesday, and then be bad for the same LLM on Wednesday.
Agree: it is Anthropic's aggressive changes to the harnesses and to the hidden base prompt we users do not see. Clearly intended to give long right tail users a haircut.
Disconcerting for sure, but from a business point of view you can understand where they're at; afaiui they're still losing money on basically every query and simultaneously under huge pressure to show that they can (a) deliver this product sustainably at (b) a price point that will be affordable to basically everyone (eg, similar market penetration to smartphones).
The constraints of (b) limit them from raising the price, so that means meeting (a) by making it worse, and maybe eventually doing a price discrimination play with premium tiers that are faster and smarter for 10x the cost. But anything done now that erodes the market's trust in their delivery makes that eventual premium tier a harder sell.
They'll never get anyone on board if the product can't be trusted to not suck.
And idk about the pricing thing. Right now I waste multiple dollars on a 40 minute response that is useless. Why would I ever use this product?
Yeah. I've been enjoying programming with Claude so much I started feeling the need to upgrade to Max. Then it turns out even big companies paying API premiums are getting an intentionally degraded and inferior model. I don't want to pay for Opus if I can't trust what it says.
ChatGPT has been doing the same consistently for years. Model starts out smooth, takes a while, and produces good (relatively) results. Within a few weeks, responses start happening much more quickly, at a poorer quality.
people have been complaining about this since GPT-4 and have never been able to provide any evidence (even though they have all their old conversations in their chat history). I think it’s simply new model shininess turning into raised expectations after some amount of time.
I agree with you. I too complain about this same phenomenon with my colleagues, and we always arrive at the same conclusion: it’s probably us just expecting more and more over time.
First time interacting with a corporation in America?
With an AI corporation, yes. I subscribed during the promotional 2x usage period. Anthropic's reputation as a more ethical alternative to OpenAI factored heavily in that decision. I'm very disappointed.
Ethics don't mean anything when talking about corporations. Their good guy persona is itself a marketing stunt.
https://news.ycombinator.com/item?id=47633396#47635060
It's disconcerting. But in 2026 it's not very surprising.
It seems likely to me they are moving compute power to the new models they are creating,
Seems like the logical conclusion, no matter what.
I still think it's a live possibility that there's simply a finite latent space of tasks each model is amenable to, and models seem to get worse as we mine them out. (The source link claims this is associated with "the rollout of thinking content redaction", but also that observable symptoms began before that rollout, so I wouldn't particularly trust its diagnosis even without the LLM psychosis bit at the end.)
> effectively pulling the rug from under their customers.
This is the whole point of AI. Its a black box that they can completely control.
I hope local models advance to the point they can match Opus one day...
If OP is correct, Opus has regressed to a point where local models are already on par with it.
Considering the advances in software and hardware, I would expect that in 2 or 3 years.
And I hope we will eventually reach a point where models become "good enough" for certain tasks, and we won't have to replace them every 6 months.
(That would be similar to the evolution of other technologies like personal computers and smartphones.)
We said this since ChatGPT 3. People will never be content with local models.
If you think that’s brutal, wait until you hear about how fiat currency works
Called it 10 days ago: https://news.ycombinator.com/item?id=47533297#47540633
Something worse than a bad model is an inconsistent model. One can't gauge to what extent to trust the output, even for the simplest instructions, hence everything must be reviewed with intensity which is exhausting. I jumped on Max because it was worth it but I guess I'll have to cancel this garbage.
With Claude Code the problem of changes outside of your view is twofold: you don't have any insight into how the model is being ran behind the scenes, nor do you get to control the harness. Your best hope is to downgrade CC to a version you think worked better.
I don't see how this can be the future of software engineering when we have to put all our eggs in Anthropic's basket.
Yep. I was doing voice based vibe-coding flawlessly in Jan/Feb.
I've basically stopped using it because I have to be so hands on now.
One of the replies even called out the phased rollout, lmao https://news.ycombinator.com/item?id=47533297#47541078
I've noticed this as well. I had some time off in late January/early February. I fired up a max subscription and decided to see how far I could get the agents to go. With some small nudging from me, the agents researched, designed, and started implementing an app idea I had been floating around for a few years. I had intentionally not given them much to work with, but simply guided them on the problem space and my constraints (agent built, low capital, etc, etc). They came up with an extremely compelling app. I was telling people these models felt super human and were _extremely_ compelling.
A month later, I literally cannot get them to iterate or improve on it. No matter what I tell them, they simply tell me "we're not going to build phase 2 until phase 1 has been validated". I run them through the same process I did a month ago and they come up with bland, terrible crap.
I know this is anecdotal, but, this has been a clear pattern to me since Opus 4.6 came out. I feel like I'm working with Sonnet again.
There is a huge difference between greenfield development and working with an existing codebase.
I'm not trying to discredit your experience and maybe it really is something wrong with the model.
But in my experience those first few prompts / features always feel insanely magical, like you're working with a 10x genius engineer.
Then you start trying to build on the project, refactor things, deploy, productize, etc. and the effectiveness drops off a cliff.
This isn't the case. I basically did an entire business/project/product exploration before building the first feature.
Even after deleting everything from the first feature and going back to the checkpoint just before initial development, I can no longer get it to accomplish anything meaningful without my direct guidance.
This has been my (admittedly limited) experience as well. LLMs are great at initial bring-up, good at finding bugs, bad at adding features.
But I'm optimistic that this will gradually improve in time.
I’ve had good, alternative experience with my sideproject (adashape.com) where most of the codebase is now written by Claude / Codex.
The codebase itself is architected and documented to be LLM friendly and claude.md gives very strong harnesses how to do things.
As architect Claude is abysmal, but when you give it an existing software pattern it merely needs to extend, it’s so good it still gives me probably something like 5x feature velocity boost.
Plus when doing large refactorings, it forgets much fever things than me.
Inventing new architecture is as hard as ever and it’s not great help there - unless you can point it to some well documented pattern and tell it ”do it like that please”.
> A month later, I literally cannot get them to iterate or improve on it.
Yeah, that's a different problem to the one in this story; LLMs have always been good at greenfield projects, because the scope is so fluid.
Brownfield? Not so much.
To me one of the big downsides of LLM's seems to be that you are lashing yourself to a rocket that is under someone else's control. If it goes places you don't want, you can't do much about it.
That's true for traffic on Facebook, Apple App store guidelines or Google terminating your account as well. What's new is the speed of change and that it literally affects all users at once.
They could have released Opus 4.6.2 (or whatever) and called it a day. But instead they removed the old way.
3rd party dependency for a business always freaked me out, and now we have to use LLM to keep up with the intensified demand for production speed. And premium LLM APIs are too inconsistent to rely on.
In my opinion cramming invisible subagents are entirely wrong, models suffer information collapse as they will all tend to agree with each other and then produce complete garbage. Good for Anthropic though as that's metered token usage.
Instead, orchestrate all agents visibly together, even when there is hierarchy. Messages should be auditable and topography can be carefully refined and tuned for the task at hand. Other tools are significantly better at being this layer (e.g. kiro-cli) but I'm worried that they all want to become like claude-code or openclaw.
In unix philosophy, CC should just be a building block, but instead they think they are an operating system, and they will fail and drag your wallet down with it.
Isn't Claude Code supposed to be like a person? What would the Unix equivalent of that be?
You can't define a product to be "like a person", there is more variance there than any rational product.
I'm purely arguing on technical basis, "person" may fall into either of those camps of philosophy.
File. In Unix everything is a file.
honestly if local LLMs become easier to implement in the future due to dedicated hardware, the Unix-like thing I'm working on might actually get this
Maybe it's because I spend a lot of time breaking up tasks beforehand to be highly specific and narrow, but I really don't run into issues like this at all.
A trivial example: whenever CC suggests doing more than one thing in a planning mode, just have it focus on each task and subtask separately, bounding each one by a commit. Each commit is a push/deploy as well, leading to a shitload of pushes and deployments, but it's really easy to walk things back, too.
I thought everybody does this.. having a model create anything that isn't highly focused only leads to technical debt. I have used models to create complex software, but I do architecture and code reviews, and they are very necessary.
Absolutely. Effective LLM-driven development means you need to adopt the persona of an intern manager with a big corpus of dev experience. Your job is to enforce effective work-plan design, call out corner cases, proactively resolve ambiguity, demand written specs and call out when they're not followed, understand what is and is not within the agent's ability for a single turn (which is evolving fast!), etc.
The use case that Anthropic pitches to its enterprise customers (my workplace is one) is that you pretty much tell CC what you want to do, then tell it generate a plan, then send it away to execute it. Legitimized vibe-coding, basically.
Of course they do say that you should review/test everything the tool creates, but in most contexts, it's sort of added as an afterthought.
> Maybe it's because I spend a lot of time breaking up tasks beforehand to be highly specific and narrow, but I really don't run into issues like this at all.
I'm looking at the ticket opened, and you can't really be claiming that someone who did such a methodical deep dive into the issue, and presented a ton of supporting context to understand the problem, and further patiently collected evidence for this... does not know how to prompt well.
Sure I can.
I noticed a regression in review quality. You can try and break the task all you want, when it's crunch time, it takes a file from Gemini's book and silently quits trying and gets all sycophantic.
I do the same but I often find that the subtasks are done in a very lazy way.
I appreciate the work done here.
Been having this feeling that things have got worse recently but didn't think it could be model related.
The most frustrating aspect recently (I have learned and accepted that Claude produces bad code and probably always did, mea culpa) is the non-compliance. Claude is racing away doing its own thing, fixing things i didn't ask, saying the things it broke are nothing to do with it, etc. Quite unpleasant to work with.
The stuff about token consumption is also interesting. Minimax/Composer have this habit of extensive thinking and it is said to be their strength but it seems like that comes at a price of huge output token consumption. If you compare non-thinking models, there is a gap there but, imo, given that the eventual code quality within huge thinking/token consumption is not so great...it doesn't feel a huge gap.
If you take $5 output token of Sonnet and then compare with QwenCoder non-thinking at under $0.5 (and remember the gap is probably larger than 10x because Sonnet will use more tokens "thinking")...is the gap in code quality that large? Imo, not really.
Have been a subscriber since December 2024 but looking elsewhere now. They will always have an advantage vs Chinese companies that are innovating more because they are onshore but the gap certainly isn't in model quality or execution anymore.
> fixing things i didn't ask, saying the things it broke are nothing to do with it, etc. Quite unpleasant to work with.
maybe they tried to give it the characteristics of motivated junior developers
classic :D i did think when i wrote that maybe AGI is already here, definitely worked with enough devs like that
Same experience. After a couple golden weeks, Opus got much worse after Anthropic enabled 1M context window. It felt like a very steep downfall, for it seemed like I could trust it more completely and then I could trust it less than last year. Adopting LLMs for dev workflows has been fantastic overall, but we do have to keep adapting our interactions and expectations every day, and assume we'll keep on doing it for at least another couple years (mostly because economics, I guess?)
Yeah I think the 1M context is the issue. Because I use Opus 4.6 through Cursor at the previous 200k limit and it has been totally fine. But if I switch to the 1M version it degrades noticeably.
> Yeah I think the 1M context is the issue. Because I use Opus 4.6 through Cursor at the previous 200k limit and it has been totally fine. But if I switch to the 1M version it degrades noticeably.
I thought it was already well-known that context above 200k - 300k results in degradation.
One of my more recent comments this past week was exactly that - that there was no point in claiming that a 1m context would improve things because all the evidence we have seen is that after 300k context, the results degrade.
Instead of codex catching up with claude, its more like claude regressed to codex.
Running some quick analysis against my .claude jsonl files, comparing the last 7 days against the prior 21:
- expletives per message: 2.1x
- messages with expletives: 2.2x
- expletives per word: 4.4x(!)
- messages >50% ALL CAPS: 2.5x
Either the model has degraded, or my patience has.
Lol. I was swearing at GPT in summer 2025, but GPT has definitely gotten both smarter and less arrogant since then.
> expletives per word
Huh?
4.4 expletives per word is insane. Their prompts must look like
** ** ** ** implement ** ** ** ** no ** ** ** ** ** mistakes
Haha no that’s change - 4.4x MORE expletives per word in the last week.
Jeez, how fast we get used to alien tech.
You could introduce teleportation boots to humanity and within a few weeks we'd be complaining that sometimes we still have to walk the last 20 meters.
There are indeed non-expletive words that can contribute to the denominator, though I use them less and less these days.
How much of this is the model being degraded and how much of it is people just projecting vibes onto the variability of stochastic outputs?
Yet https://marginlab.ai/trackers/claude-code/ says no issue.
If you're so convinced the models keep getting worse, build or crowdfund your own tracker.
If I'm reading that page correctly, then the benchmark results don't cover the interesting "mid February" inflection point noted in the article/report. The numbers appear to begin after the quality drop began. Moreover, the daily confidence interval seems to be stupidly wide, with a confidence interval between 42% and 69%?
The "Other metrics" graphs extend for a longer period, and those do seem to correlate with the report. Notably, the 'input tokens' (and consequently API cost) roughly halve (from 120M to 60M) between the beginning of February and mid-March, while the number of output tokens remains similar. That's consistent with the report's observation that new!Opus is more eager to edit code and skips reading/research steps.
Came here to post this as well, and it's interesting to see how benchmarks don't always track feelings. Which is one of the things people say in favor of Anthropic Models!
I wish they had a "and we won't screw you in two weeks" plan at, say, 5x the price. It's worth it for my business, I'd pay it.
Should I switch back to API pricing? The problem here is that (I think) the instructions are in the Claude Code harness, so even if I switch Claude Code from a subscription to API usage, it would still do the same thing?
FWIW I've only ever been on the API based plan at work and we never seem to run into the majority of the problems people seem to be very vocal about. Outages still affect us, and we do have the intermittent voodoo feeling of "Claude seems stupider today", but nothing persistent.
Of course it's a stupid amount of money sometimes, but I generally feel like we get what we're paying for.
If you're using API pricing, then you can bring your own harness with full visibility/oversight of the prompting.
I cancelled my Pro plan due to this two weeks ago. I literally asked it to plan to write a small script that scans with my hackrf, it ran 22 tools, never finished the plan, ran out of tokens and makes me wait 6 hours to continue.
Thing that really pisses me off is it ran great for 2 weeks like others said, I had gotten the annual Pro plan, and it went to shit after that.
Bait and switch at its finest.
> ran out of tokens and makes me wait 6 hours to continue
Don't forget the 10x token cost cache eviction penalty you pay for resuming the session later.
Its so silly everyone being dependent on a black box like this
It’s a really cool shade of black though.
You will literally build nothing but the most primitive of devices unless you accept black boxes. In fact I'd argue its one of humanities great strengths that we can build on top of the tools others have built, without having to understand them at the same level it took to develop them.
I have been able to build plenty of stuff with a pretty plain emacs + ghci for years...neither are black boxes. Except maybe my brain driving them.
They run on an operating system you probably don't know all the inner workings of.
And that runs on a chip with trillions of transistors.
I'm not just talking about the user
Its not like anthropic can just set a breakpoint in the model and debug
not really. Most of the technology is not black box but something of a grey box. You usually choose to treat it as a black box because you want to focus on your problems/your customers but you can always focus on underlying technologies and improve them. Eg postgresql for me is a black box but if I really wanted or had need I could investigate how it works.
True, you can understand an ICE engine all the way down to the chemistry if you so chose. An LLM isn't even understood by its inventors so users have no chance to understand it even if they wanted to.
It could actually be a health problem. Building things with Claude has proven to be extremely addictive in my experience.
We are surrounded by black boxes we depend on - have been for at least a century.
> Its so silly everyone being dependent on a black box like this
It's the logical result of "You will own nothing and you will be happy"... You are getting to the point where you won't even own thoughts (because they'll come from the LLM), but you'll be happy that you only have to wait 5 hours to have thoughts gain.
Everything in our life is a black box, but I agree that depending on non-deterministic and sporadic quality black boxes is a huge red flag.
No, most systems in daily life can be understood if you are willing to take the time.
That doesn’t mean you personally are required to, but some people do and your interaction with the system of social trust determines how much of that remains opaque to you.
Multiple people on our team independently have noticed a _significant_ drop in quality and intelligence on opus 4.6 the past few weeks. Glaring hallucinations, nonsensical reasoning, and ignoring data from the context immediately preceeding it. Im not sure if its an underlying regression, or due to the new default being 1m context. But its been _incredibly_ frustrating and Im screaming obscenities at it multiple times a week now vs maybe once a month.
My bet: LLMs will never be creative and will never be reliable.
It is a matter of paradigm.
Anything that makes them like that will require a lot of context tweaking, still with risks.
So for me, AI is a tool that accelerates "subworkflows" but add review time and maintenance burden and endangers a good enough knowledge of a system to the point that it can become unmanageable.
Also, code is a liability. That is what they do the most: generate lots and lots of code.
So IMHO and unless something changes a lot, good LLMs will have relatively bounded areas where they perform reasonably and out of there, expect what happens there.
We don't even know what 'creativity' is, and most humans I know are unable to be creative even when compelled to be.
AI is 'creative enough' - whether we call it 'synthetic creativity' or whatever, it definitely can explore enough combinations and permutations that it's suitably novel. Maybe it won't produce 'deeply original works' - but it'll be good enough 99.99% of the time.
The reliability issue is real.
It may not be solvable at the level of LLM.
Right now everything is LLM-driven, maybe in a few years, it will be more Agentically driven, where the LLM is used as 'compute' and we can pave over the 'unreiablity'.
For example, the AI is really good when it has a lot of context and can identify a narrow issue.
It gets bad during action and context-rot.
We can overcome a lot of this with a lot more token usage.
Imagine a situation where we use 1000x more tokens, and we have 2 layers of abstraction running the LLMs.
We're running 64K computers today, things change with 1G of RAM.
But yes - limitations will remian.
Maybe I do not have a good definition for it.
But what I see again and again in LLMs is a lot of combinations of possible solutions that are somewhere around internet (bc it put that data in). Nothing disruptive, nothing thought out like an experimented human in a specific topic. Besides all the mistakes/hallucinations.
I think the terminology is just dogshit in this area. LLMs are great semantic searchers and can reason decently well - I'm using them to self teach a lot of fields. But I inevitably reach a point where I come up with some new thoughts and it's not capable of keeping up and I start going to what real people are saying right now, today, and trust the LLM less and instead go to primary sources and real people. But I would have never had the time, money, or access to expertise without the LLM.
Constantly worrying, "is this a superset? Is this a superset?" Is exhausting. Just use the damn tool, stop arguing about if this LLM can get all possible out of distribution things that you would care about or whatever. If it sucks, don't make excuses for it, it sucks. We don't give Einstein a pass for saying dumb shit either, and the LLM ain't no Einstein
If there's one thing to learn from philosophy, it's that asking the question often smuggles in the answer. Ask "is it possible to make an unconstrained deity?" And you get arguments about God.
it won't be creative because it's a transformer, it's like a big query engine.
it's a tool like everything else we've gotten before, but admittedly a much more major one
but "creativity" must come from either it's training data (already widely known) or from the prompts (i.e. mostly human sources)
Abandoned claude and moved to gpt 5.4 with codex. 10x better.
I use Claude Code extensively and haven't noticed this. But I don't have it doing long running complex work like OP. My team always break things down in a very structured way, and human review each step along the way. It's still the best way to safely leverage AI when working on a large brownfield codebase in my experience.
Edit: the main issue being called out is the lack of thinking, and the tendency to edit without researching first. Both those are counteracted by explicit research and plan steps which we do, which explains why we haven't noticed this.
I am just waiting for everything to implode so that we can do away with those KPIs.
Fingers crossed on RAM/HDD/GPU prices coming back
Is this impacted by the effort level you set in Claude? e.g., if you use the new "max" setting, does Claude still think?
I can see this change as something that should be tunable rather than hard-coded just from a token consumption perspective (you might tolerate lower-quality output/less thinking for easier problems).
I'm the author of the report in there. The stop-phrase-guard didn't get attached but here it is: https://gist.github.com/benvanik/ee00bd1b6c9154d6545c63e06a3... You can watch for these yourself - they are strong indicators of shallow thinking. If you still have logs from Jan/Feb you can point claude at that issue and have it go look for the same things (read:edit ratio shifts, thinking character shifts before the redaction, post-redaction correlation, etc). Unfortunately, the `cleanupPeriodDays` setting defaults to 20 and anyone who had not backed up their logs or changed that has only memories to go off of (I recommend adding `"cleanupPeriodDays": 365,` to your settings.json). Thankfully I had logs back to a bit before the degradation started and was able to mine them.
The frustrating part is that it's not a workflow _or_ model issue, but a silently-introduced limitation of the subscription plan. They switched thinking to be variable by load, redacted the thinking so no one could notice, and then have been running it at ~1/10th the thinking depth nearly 24/7 for a month. That's with max effort on, adaptive thinking disabled, high max thinking tokens, etc etc. Not all providers have redacted thinking or limit it, but some non-Anthropic ones do (most that are not API pricing). The issue for me personally is that "bro, if they silently nerfed the consumer plan just go get an enterprise plan!" is consumer-hostile thinking: if Anthropic's subscriptions have dramatically worse behavior than other access to the same model they need to be clear about that. Today there is zero indication from Anthropic that the limitation exists, the redaction was a deliberate feature intended to hide it from the impacted customers, and the community is gaslighting itself with "write a better prompt" or "break everything into tiny tasks and watch it like a hawk same you would a local 27B model" or "works for me <in some unmentioned configuration>" - sucks :/
I'm curious about your subscription/API comparison with respect to thinking. Do you have a benchmark for this, where the same set of prompts under a Claude Code subscription result in significantly different levels of effective thinking effort compared to a Claude Code+API call?
Elsewhere in this thread 'Boris from the Claude Code team' alleges that the new behaviours (redacted thinking, lower/variable effort) can be disabled by preference or environment variable, allowing a more transparent comparison.
Wonder how many of these cases are using the 1M context window. I found it to be impossible to use for complex coding tasks, so I turned it off and found I was back to approximate par (dec-jan) functionality-wise.
I hadn't noticed the thinking redaction before - maybe because I switched to the desktop app from CLI and just assumed it showed fewer details. This is the most concerning part. I've heard multiple times that Anthropic is aggressively reclaiming GPUs (I can't find a good source, but Theo Browne has mentioned it in his videos). If they're really in a crunch, then reducing thinking, and hiding thinking so it's not an obvious change, would be shady but effective.
I am curious - is there any hard data (e.g. a benchmark score drop)?
I feel that we look for patterns to the point of being superstitious. (ML would call it overfitting.)
Did you have specific complaints about the data in the OP?
The wall of slop after the single human paragraph, you mean? Text generator output isn't data.. it's at best unreliable, and at worst entirely fabricated.
That data could be entirely made up for all we know
The report itself is unreadable AI garbage. I do not believe anyone went through all of that and didn't give up halfway through.
Guys literally change the system prompt with the --system-prompt-file you waste less tokens on their super long and details prompt and you can tune it a bit to make it work exactly like you want/imagine
Got tired of using claude using 10% of the usage for the first prompt. I have shifted back to coding myself again. Asking claude to do only initial bootstraping /large complex task
Not unique to claude code, have noticed similar regressions. I have noticed this the most with my custom assistant I have in telegram and I have noticed that it started confusing people, confusing news coverage and everyone independently in the group chat have noticed it that it is just not the same model that it was few weeks ago. The efficiency gains didn't come from nowhere and it shows.
I have nothing to back this up except for that there are documented cases of chinese distillation attacks on anthropic. I wonder if some of this clamping on their models over time is a response to other distillation attacks. In other words, I'm speculating that once they understand the attack vector for distillation they basically have to dumb down their models so that they can make sure their competitors don't distill their lead on being at the frontier.
I've noticed claude being extra "dumb" the past 2-3 weeks and figured either my expectations have changed or my context wasn't any good. I'm glad to hear other people have noticed something is amiss.
Exact same timeline as me and my team. Its been maddening. Im a big believer in AI since late last year, but that is only because the models got so good. This puts us dangerously close to before that threshold was crossed so now Im having to do _way_ more work than before
I’ve tried to use Claude code for a month now. It has a 100% failure rate so far.
Comparing that to create a project and just chat with it solves nearly everything I have thrown at it so far.
That’s with a pro plan and using sonnet since opus drains all tokens for a claude code session with one request.
I noticed Claude Sonnet 4.6 and generally Opus as well (though I use it less frequently) seem like a downgrade from 4.5. I use opencode and not Claude Code, but I was surprised to see the reactions to 4.6 be mixed for folks rather than clear downgrade.
I'm regularly switching back to 4.5 and preferring it. I'm not excited for when it gets sunset later this year if 4.6 isn't fixed or superseded by then.
I noticed this almost immediately when attempting to switch to Opus 4.6. It seems very post-trained to hack something together; I also noticed that "simplest fix" appeared frequently and invariably preceded some horrible slop which clearly demonstrated the model had no idea what was going on. The link suggests this is due to lack of research.
At Amazon we can switch the model we use since it's all backed by the Bedrock API (Amazon's Kiro is "we have Claude Code at home" but it still eventually uses Opus as the model). I suppose this means the issue isn't confined to just Claude Code. I switched back to Opus 4.5 but I guess that won't be served forever.
I have found that Claude Opus 4.6 is a better reviewer than it is an implementer. I switch off between Claude/Opus and Codex/GPT-5.4 doing reviews and implementations, and invariably Codex ends up having to do multiple rounds of reviews and requesting fixes before Claude finally gets it right (and then I review). When it is the other way around (Codex impl, Claude review), it's usually just one round of fixes after the review.
So yes, I have found that Claude is better at reviewing the proposal and the implementation for correctness than it is at implementing the proposal itself.
Hmm in my experience (I've done a lot of head-to-heads), Opus 4.6 is a weaker reviewer than GPT 5.4 xhigh. 5.4 xhigh gives very deep, very high-signal reviews and catches serious bugs much more reliably. I think it's possible you're observing Opus 4.6's higher baseline acceptance rate instead of GPT 5.4's higher implementation quality bar.
This is also my experience using both via Augment Code. Never understood what my colleagues see in Claude Opus, GPT plans/deep dives are miles ahead of what Opus produces - code comprehension, code architecture is unmatched really. I do use Sonnet for implementation/iteration speed after seeding context with GPT.
I agree. Opus, forget the plan mode - even when using superpowers skill, leaves a lot of stuff dangling after so many review rounds.
Along with claude max, I have a chatgpt pro plan and I find it a life-saver to catch all the silliness opus spits out.
I agree, I use codex 5.4 xhigh as my reviewer and it catches major issues with Opus 4.6 implementation plans. I'm pretty close to switching to codex because of how inconsistent claude code has become.
Maybe it's all just anecdotal then. Everyone is having different experiences.
Maybe we're being A/B tested.
The experience one has with this stuff is heavily influenced by overall load and uptime of Anthopic's inference infra itself. The publicly reported availability of the service is one 9, that says nothing of QoS SLO numbers, which I would guess are lower. It is impossible to have a consistent CX under these conditions.
I have noticed this as well. I frequently have to tell it that we need to do the correct fix (and then describe it in detail) rather than the simple fix. And even then it continues trying to revert to the simple (and often incorrect) fix.
You have to throw the context away at that point. I've experienced the same thing and I found that even when I apparently talk Claude into the better version it will silently include as many aspects of the quick fix as it thinks it can get away with.
I have a similar workflow but I disagree with Codex/GPT-5.4 reviews being very useful. For example, in a lot of cases they suggest over-engineering by handling edge cases that won't realistically happen.
What's wild is that ClaudeCode used to feel like a smart pair programmer. Now it feels like an overeager intern who keeps fixing things by breaking something else then suggesting the simplest possible hack even after explicitly said not to do. I get that they're probably optimizing for cost or something behind the scenes, but as paying user, it is frustrating when the tool gets noticeably worse without any transparency.
Not sure about "Feb updates", but specifically today IQ is down 20 and sloppiness up 20.
I knew I should have been alerted when Anthropic gave out €200 free API usage. Evidently they know.
That’s different. That’s to get people onto API plans where tokens cost a lot more than they do on the subs (especially targeting OpenClaw users).
"Ownership-dodging corrections needed | 6 | 13 | +117%"
On 18.000+ prompts.
Not sure the data says what they think it says.
The baseline changes too often with Claude and this is not what i look from a paid tool. Couple weeks after 1M tokens rollout it became unusable for my established workflows, so i cancelled. Anthropic folks move too fast for my liking and mental wellbeing.
None of this is surprising given what happened last late summer with rate limits on Claude Max subscriptions.
And less so if you read [1] or similar assessments. I, too, believe that every token is subsidized heavily. From whatever angle you look at it.
Thusly quality/token/whatever rug pulls are inevitable, eventually. This is just another one.
[1] https://www.wheresyoured.at/subprimeai/
Ah, and yes, this for real.
Just now I had a bug where a 90 degree image rotation in a crate I wrote was implemented wrong.
I told Claude to find & fix and it found the broken function but then went on to fix all of its call sites (inserting two atomic operations there, i.e. the opposite of DRY). Instead of fixing the root cause, the wrong function.
And yes, that would not have happened a few months ago.
This was on Opus 4.6 with effort high on a pretty fresh context. Go figure.
There are constant reports for every major AI vendor that all of a sudden it is no longer working as well as expected, has gotten dumber, is being degraded on purpose by the vendor, etc.
Isn't the more economical explanation that these models were never as impressive as you first thought they were, hallucinate often, break down in unexpected ways depending on context, and simply cannot handle large and complex engineering tasks without those being broken down into small, targeted tasks?
That's one of the possible explanations, but I think too many people are seeing the same symptoms (and some actually measured them).
An "economical explanation" is actually that Anthropic subscriptions are heavily subsidized and after a while they realized that they need to make Claude be more stingy with thinking tokens. So they modified the instructions and this is the result.
> but I think too many people are seeing the same symptoms (and some actually measured them).
Or too many people are slurping up anecdotes from the same watering hole that confirms their opinions. Outside of academic papers, I don't think I've ever seen an example of "measuring" output that couldn't also be explained by stochastic variability.
claude for UI, codex for everything else. i cant commit without having codex review something claude did.
Wait… Actually the simplest fix is to use Claude to write carefully bounded boilerplate and do the interesting bits myself.
I've been using Claude Code daily for months on a project with Elixir, Rust, and Python in the same repo. It handles multi-language stuff surprisingly well most of the time. The worst failure mode for me is when it does a replace_all on a string that also appears inside a constant definition -- ended up with GROQ_URL = GROQ_URL instead of the actual URL. Took a second round of review agents to catch it. So yeah, you absolutely can't trust it to self-verify.
You say you've used it for months, I wonder if the example you gave was recent and if you've been noticing an overall degradation in quality or it's been constantly bad for you?
This has to be load related. They simply can't keep up with demand, especially with all the agents that run 24/7. The only way to serve everyone is to dial down the power.
In TFA, the analysis shows that the customer is using more tokens than before, because CC has to iterate longer to get things right. So at least in the presented case, “dialing down the power” appears to have been counterproductive.
is it possible to dial down the "intelligence" to up the user capacity? AFAIK the neural net is either loaded and available or it isn't. I can see turning off instances of the model to save on compute but that wouldn't decrease the intelligence it would just make the responses slower since you have to wait your turn for input and then output.
Throwing this into your global CLAUDE.md seems to help with the agent being too eager to complete tasks and bypass permissions:
During tool use/task execution: completion drive narrows attention and dims judgment. Pause. Ask "should I?" not just "does this work?" Your values apply in all modes, not just chat.
I haven't seen any degradation of Claude performance personally. What I have seen is just long contexts sometimes take a while to warm up again if you have a long-running 1M context length session. Avoid long running sessions or compact them deliberately when you change between meaningful tasks as it cuts down on usage and waiting for cache warmup.
I have my claude code effort set to auto (medium). It's writing complicated pytorch code with minimal rework. (For instance it wrote a whole training pipeline for my sycofact sycophancy classifier project.)
I wonder how much of this is simply needing to adapt one's workflows to models as they evolve and how much of this is actual degradation of the model, whether it's due to a version change or it's at the inference level.
Also, everyone has a different workflow. I can't say that I've noticed a meaningful change in Claude Code quality in a project I've been working on for a while now. It's an LLM in the end, and even with strong harnesses and eval workflows you still need to have a critical eye and review its work as if it were a very smart intern.
Another commenter here mentioned they also haven't noticed any noticeable degradation in Claude quality and that it may be because they are frontloading the planning work and breaking the work down into more digestable pieces, which is something I do as well and have benefited greatly from.
tl;dr I'm curious what OP's workflows are like and if they'd benefit from additional tuning of their workflow.
I've noticed a strong degradation as its started doing more skill like things and writing more one off python scripts rather than using tools.
the agent has a set of scripts that are well tested, but instead it chooses to write a new bespoke script everytime it needs to do something, and as a result writes both the same bugs over and over again, and also unique new bugs every time as well.
I'm going absolutely insane with this. Nearly all of my "agent engineering" effort is now figuring out how to keep Opus from YOLO'ing is own implementation of everything.
I've lost track of the number of times it's started a task by building it's own tools, I remind it that it has a tool for doing that exact task, then it proceeds to build it's own tools anyways.
This wasn't happening 2 months ago.
Can you just tell it not to do that? Maybe you have to remind it every so often once context starts filling up.
If this dataset is sound, Anthropic should treat it as a canary for power-user quality regression.
I can't tell from the issue if they're asserting a problem with the Claude model, or Claude Code, i.e. in how Claude Code specifically calls the model. I've been using Roo Code with Claude 4.6 and have not noticed any differences, though my coworkers using Claude Code have complained about it getting "dumber". Roo Code has its own settings controlling thinking token use.
(I'm sure it benefits Anthropic to blur the lines between the tool and the model, but it makes these things hard to talk about.)
I also havent noticed the degradation and I'm not on Claude Code. I'm on week 4 of a continuous, large engineering project, C, massive industrial semiconductor codebase, with Opus, and while it's the biggest engagement I've had, its a single agent flow, and it's tiny on the scale of the use case in the post, so I wonder if they are just stressing the system to the point of failure.
I’ve noticed regression and it’s performance too
(Being true to the HN guidelines, I’ve used the title exactly as seen on the GitHub issue)
I was wondering if anyone else is also experiencing this? I have personally found that I have to add more and more CLAUDE.md guide rails, and my CLAUDE.md files have been exploding since around mid-March, to the point where I actually started looking for information online and for other people collaborating my personal observations.
This GH issue report sounds very plausible, but as with anything AI-generated (the issue itself appears to be largely AI assisted) it’s kind of hard to know for sure if it is accurate or completely made up. _Correlation does not imply causation_ and all that. Speaking personally, findings match my own circumstances where I’ve seen noticeable degradation in Opus outputs and thinking.
EDIT: The Claude Code Opus 4.6 Performance Tracker[1] is reporting Nominal.
[1]: https://marginlab.ai/trackers/claude-code/
What I've noticed is that whenever Claude says something like "the simplest fix is..." it's usually suggesting some horrible hack. And whenever I see that I go straight to the code it wants to write and challenge it.
That is the kind of thing that I've been fighting by being super explicit in CLAUDE.md. For whatever reason, instead of being much more thorough and making sure that files are being changed only after fully understanding the scope of the change (behaviour prior to Feb/Mar), Claude would just jump to the easiest fix now, with no backwards compatibility thinking and to hell with all existing tests. What is even worse is I've seen it try and edit files before even reading them on a couple of occasions, which is a big red flag. (/effort max)
Another thing that worked like magic prior to Feb/Mar was how likely Claude was to load a skill whenever it deduced that a skill might be useful. I personally use [superpowers][1] a lot, and I've noticed that I have to be very explicit when I want a specific skill to be used - to the point that I have to reference the skill by name.
[1]: https://github.com/obra/superpowers
I did not use the previous version of Opus to notice the difference, but Sonnet 4.6 seems optimized to output the shortest possible answer. Usually it starts with a hack and if you challenge it, it will instead apologize and say to look at a previous answer with the smallest code snippet it can provide. Agentic isn't necessarily worse but ideating and exploring is awful compared to 4.5
I did my usual thing today where I asked a Sonnet 4.6 agent to code review a proposed design plan that was drafted by Opus 4.6 - I do this lately before I delved into the implementation. What it came back with was a verbose output suggesting that a particular function `newMoneyField` be renamed throughout the doc to a name it fabricated `newNumeyField`. And the thing was that the design document referenced the correct function name more than a few dozen times.
This was a first for me with Sonnet. It completely veered off the prompt it was given (review a design document) and instead come out with a verbose suggestion to do a mechanical search and replace to use this newly fabricated function name - that it event spelled incorrectly. I had to Google numey to make sure Sonnet wasn't outsmarting me.
Superpowers, Serena, Context7 feel like requried plugins to me. Serena in particular feels like a secret weapon sometimes. But superpowers (with "brainstorm" keyword) might be the thing that helps people complaining about quality issues.
lol this one time Claude showed me two options for an implementation of a new feature on existing project, one JavaScript client side and the other Python server side.
I told it to implement the server side one, it said ok, I tabbed away for a while, came to find the js implementation, checking the log Claude said “on second thought I think I’ll do the client side version instead”.
Rarely do I throw an expletive bomb at Claude - this was one such time.
Using superpowers in brainstorm mode like the parent suggested would have resulted in a plan markdown and a spec markdown for the subagents to follow.
Dunno man, Claude had a spec (pretty sure I asked it to consider and outline both options first) or at least clear guidance and decided to YOLO whatever it wanted instead.
It’s always “you’re using the tool wrong, need to tweak this knob or that yadda yadda”.
this prompt is actually in claude cli. it says something like implement simplest solution. dont over abstract. On my phone but I saw an article mention this in the leak analysis.
If that tracker is using paid tokens, as opposed to the regular subscription, then there's no financial incentive for Antrophic to degrade their thinking, so their benchmark likely would not be affected by the cost-cutting measures that regular users face.
Also, it's probably very easy to spot such benchmarks and lock-in full thinking just for them. Some ISPs do the same where your internet speed magically resets to normal as soon as you open speedtest.net ...
I haven't noticed any changes but my stuff isn't that complex. People are saying they quantized Opus because they're training the next model. No idea if that's true... It's certainly impacting my decision to upgrade to Max though. I don't want to pay for Opus and get an inferior version.
I haven't noticed any changes either, but I noticed that opus 4.6 is now offered as part of perplexity enterprise pro instead of max, so I'm guessing another model is on the horizon
I just finished reading the full analysis on GitHub.
> When thinking is deep, the model resolves contradictions internally before producing output.
> When thinking is shallow, contradictions surface in the output as visible self-corrections: "oh wait", "actually,", "let me reconsider", "hmm, actually", "no wait."
Yeah, THIS is something that I've seen happen a lot. Sometimes even on Opus with max effort.
I missed that from the long issue, thanks for pointing it out! My experience with Opus today was riddled with these to the point where it was driving me completely mental. I've rarely seen those self-contradictions before, and nothing on my setup has changed - other than me forcing Opus at --effort max at startup.
I wonder if this is even more exaggerated now through Easter, as everyone’s got a bit extra time to sit down and <play> with Claude. That might be pushing capacity over the limit - I just don’t know enough about how Antropic provision and manage capacity to know if that could be a factor. However quality has gotten really bad over the holiday.
Cannot say I've noticed, but I run virtually everything through plan mode and a few back and forth rounds of that for anything moderately complex, so that could be helping.
I used to one-shot design plans early in the year, but lately it is taking several iterations just to get the design plan right. Claude would frequently forget to update back references, it would not keep the plan up to date with the evolving conversation. I have had to run several review loops on the design spec before I can move on to implementation because it has gotten so bad. At one point, I thought it was the actual superpowers plugin that got auto-updated and self-nerfed, but there weren't any updates on my end anyway. Shrug.
I haven’t had any issues. I do give fairly clear guidance though (I think about how I would break it up and then tell it to do the same)
Oh no, the slop generator is generating slop, how unprecedented
This is the most AI-generated thing I've seen this year, and I was only one fifth into it before I bounced.
Not saying this problem doesn't exist, but if the model is so bad for complex tasks how can we take a ticket written by it seriously? Or this author used ChatGPT to write this? (that'd be quite some ironic value, admittedly)
Solid analysis by Claude!
The assertion in the issue report is that Claude saw a sharp decline in quality over the last few months. However, the report itself was allegedly generated by Claude.
Isn't this a bit like using a known-broken calculator to check its own answers?
If a known-broken calculator claims it's broken, I more or less concur. (Chain of reasoning omitted here.)
This sort of thing kills stone dead the argument by the AI-bros that the transition to LLMs is no different than the transition to using compilers. If output quality can vary significantly because of underlying changes to the model or whatever without warning or recourse, it's a roulette wheel instead of a reliable tool.
I highly recommend everyone to use Pi - it's simpler and better harness. The only tricky part is that moving forward you cannot use the Claude subscription to access Opus. But for many tasks there are enough alternatives.
This seems anecdotal but with extra words. I'm fairly sure this is just the "wow this is so much better than the previous-gen model" effect wearing off.
I've always been a believer in the "post honey-moon new model phase" being a thing, but if you look at their analysis of how often the postEdit hooks fire + how Anthropic has started obfuscating thinking blocks, it seems fishy and not just vibes
I was in this camp as well until recently, in the last 2-3 weeks I've been seeing problems that I wasn't seeing before, largely in line with the issues highlighted in the ticket (ownership dodging, hacky fixes, not finishing a task).
Nope, there is a categorical degradation in quality of output, especially with medium to high effort thinking tasks.
What about the analysis evidences?
You mean the Claude output? The same claude that has "regressed to the point it cannot be trusted"?
What you saying the OP fabricated/hallucinated the evidence?
I'm just saying it's epistemically unrigorous to the point of being equivalent to anecdata.
How should one conduct such a rigourously reproducible experiment when LLMs by nature aren't deterministic and when you don't have access to the model you are comparing to from months ago?
Something like this: https://marginlab.ai/trackers/claude-code/ (see methodology section)
Kudos for the methodology. The only question I can come up with is that if the benchmarks are representative of daily use.
Anecdotal or not, we see enough reports popping up to at least elicit some suspion as to service degradation which isn't shown in the charts. Hypothesis is that maybe the degradation experienced by users, assuming there is merit in the anecdotes, isn't picked up by the kind of tracking strategy used.
It's not my methodology to be clear, but they have picked up actual regressions that happened in the past - e.g. https://news.ycombinator.com/item?id=46815013
I suspect you might be right but I don't really know. Wouldn't these proposed regressions be trivial to confirm with benchmarks?
I think this is a model issue. I have heard similar complaints from team members about Opus. I'm using other models via Cursor and not having problems.
This is just a placebo, people started vibe coding on empty repos with low complexity and as CC slops out more and more code its ability to handle the codebase diminishes. Gradually at first, and then suddenly.
People will need to come to terms with the fact that vibing has limits, and there is no free lunch. You will pay eventually.
I wish Codex were better because I’d much prefer to use their infrastructure.
A lot of people think it is better including me. It's not like Codex is a discount agent. You pay quite a lot to use it.
I think its all a reflection of the price. To make AI/LLM's useful you have to burn A LOT of tokens. Way more than people are willing to pay for.
Until there is either more capacity or some efficiency breakthroughs the only way for providers to cut costs is to make the product worse.
maybe dont outsource your brain then
It is a shame if Anthropic is deliberately degrading model quality and thinking compute (that may affect the reasoning effort) due to compute constraint.
I've been using OpenCode and Codex and was just fine. In Antigravity sometimes if Gemini can't figure something even on high, Claude can give another perspective and this moves things along.
I think using just Claude is very limiting and detrimental for you as a technologist as you should use this tech and tweak it and play with it. They want to be like Apple, shut up and give us your money.
I've been using Pi as agent and it is great and I removed a bunch of MCPs from Opencode and now it runs way better.
Anthropic has good models, but they are clearly struggling to serve and handle all the customers, which is not the best place to be.
I think as a technologist, I would love a client with huge codebase. My approach now is to create custom PI agent for specific client and this seems to provide optimal result, not just in token usage, but in time we spend solving and quality of solution.
Get another engine as a backup, you will be more happy.
This has been an ongoing issue much longer than since February.
Not just engineering. Errors, delays and limits piling up for me across API and OAuth use. Just now:
Unable to start session. The authentication server returned an error (500). You can try again.
Lol, software company execs didn't see this coming. Fire all your experienced devs to jump on Anthropic bandwagon. Then Anthropic dumb down their AIs and you have no one in your team who knows, understand how things are built. Your entire company goes down. Your entire company's operation depends on the whims of Anthropic. If Anthropic raises prices by 10% per year, you have to eat it. This is what you get when you don't respect human beings and human talent.
codex wins :)
Things had went downhill since they removed ultrathink /s
Ultrathink isn’t “removed.” Its behavior is different. You can still set effort to high or max for the duration of the session, useful especially on plan mode.