Last weekend I bought my wife a bike off marketplace. It was in good condition but was missing one of the internal cable routing grommets. I gave Claude pictures of the pill-shaped hole by itself and with my digital calipers in the long and short directions.
Gave it a short prompt and it gave me an openscad model with everything parametrized. I printed with no changes in tpu and it was nearly perfect on the first try. Claude put in a 0.3mm subtraction in the x/y dimensions and I lowered it to 0.1 and it's perfect.
Much easier shape than ancient Roman architecture but still very cool how easy it was.
Yeah, CAD has been my personal example of "oh the barrier to entry for this skill was high enough that I didn't do it and now I can be passably bad at it enough to get some simple things done"
I've had similar experiences with making simple functional parts off a 3d printer with OpenSCAD + LLMs. I'm very aware that the models are worse at it than say, generating react code, and I'm also the antithesis of a skilled pilot. It's still cool and has resulted in me starting to learn a new skill at a hobby level.
It's like this with a lot of things now. For example, Nix's learning curve used to be a huge barrier to entry. Now with LLMs, I'm using nix-darwin and home-manager for dotfiles, package management, and have individual flakes in all of my projects for cryptographically reproducible builds!
Claude does well if you can provide all dimensions. It fails at guessing though. The real magic is when you can provide one dimension or photograph with a ruler in it and the AI will figure the rest out. Right now, Claude anyways, is pretty bad at guessing.
I was recently trying to get models to generate a 3D fortune cookie. Claude in three.js and Gemini in openSCAD. Neither really got the concept or could get very close at all. It's a surprisingly complex shape I guess.
You optimize for no support when selecting print orientation (but for anything semi-cylindrical like described that would be the only sane orientation and the one slicer would choose when you smash the 'Auto Orientation' button).
> Antigravity was the only autonomous agent that implemented the Pantheon’s signature interior ceiling pattern: repeated square coffers visible through the oculus.
That is seriously really impressive. I looked at the 3D model and didn't even thing to LOOK INSIDE the building before reading this.
Here's [1] the 3D model with `show_cutaway` enabled.
Antigravity may well Top the whatever benchmark but:
My Antigravity (forced) replacement for Gemini CLI requires me to log on via browser every time I use it, and my Antigravity IDE won't update at all, so:
If it's ok I'd prefer they just work on reaching a baseline acceptable rollout before worrying about being Top in anything.
I agree, my main concern regarding Google AI products is this endless pain around the UX of login / billing / upgrades / product sunsets... but their LLM models are good and Antigravity 2.0 is not that bad either (unless you lost all you Antigravity 1.0 setup and projects - like many people did)
I just left the google I/O feeling less confident about google's execution here.
- Gemini 3.5 flash is strange. Old cutoff, basically better than 3.1 pro at soem things worse at others, sometimes cheaper, sometimes more expensive than 3.1 pro.
- Antigravity had seemed abandoned, and people speculated them cutting it off, and they kind of did migrating everyone to a new antigravity
- Google "shipped the org chart" and they have so many AI products and none seem best of breed (e.g. the Gemini integration in google docs is worse than claude)
I was actually hoping for "Opus level intelligence at Haiku costs" model or "Sonnet level performance in Gemini 3.0 pricing", either of these would have been a workhorse, plus a competitor to Claude/Codex (1 app to do things). I got neither.
I just use Claude Code and intellij, so I don't understand why so many people complain about Antigravity ditching VS Code, what's the surface not covered by using Antigravity CLI + VS Code (or any other IDE)?
Gemini cli was open source. Antigravity cli is not. Not at feature parity, missing many features and now we are forced to migrate away from Gemini cli before anti gravity cli is ready.
The difference in its ability is immense. Even with less features it makes a lot of sense to switch. It really shows how much the harness matters almost equally to the model.
At least one of the missing features is a basic piece of functionality (showing token quota used). Without it, you're pretty much guaranteed to get locked out for a week with no warning.
I'm not GP, but I am somewhat excited about antigravity CLI. I adopted Gemini CLI early and really liked it, though over time it got dumber and dumber until a point when I realized it was foolish to use it instead of claude/codex. I'm hopefuly that antigravity CLI won't go through that path, but also can't fight a skepticism.
I don’t think it’s the cli that was dumber, just the model it was using. They drastically reduced limits on their best model so that’s likely how you got stuck downgrading model and getting worse results.
I'm sensing in reality that behind the scenes there is a difficult trade-off between quantization and usage limits. You can have a "smart" model but poor limits, or good limits and a "dumb" model.
This seems very similar to mobile data limits (remember those years?), where there wasn't enough tower bandwidth to serve everyone unlimited data, so telecos were in constant tension between data caps and bandwidth throttling.
It wasn't until 5G came along with 100x network capacity that they could finally give everyone "unlimited" data.
I've got an AI pro plan and haven't been able to log in for months. Endless checking in with my google support guy. At least Dinesh wishes me good health every week, so that's nice.
Having my workflow disrupted is the main reason I never adopted Antigravity, despite liking it. I'm glad to see G is invested, but the older I get the more protective I am of my workflow.
Exactly. I admit it's a bit extreme, but this is a big reason why I insist that neovim is my IDE, and I won't adopt anything else. If I can't make it work in neovim, I will move to something else (unless I have no choice, but that happens very rarely at this point).
The forced upgrade from Gemini CLI which I liked as much, and as some ways better than Claude Code was bad. But them just sending out that email on Wednesday that basically said "Thanks for subscribing to Google One AI Pro, as of right now we're adding limits to your account. Tough shit you get nothing." left a REALLY bad taste in my mouth. I had previously praised the "AI Pro" subscription as a good value.
I quit AI Pro earlier this year for the same reason. I went to use it one day (I don't think I'd even used it much in the preceding week) and found that my limits had been reduced overnight and my usage was already too high. I had something like a 7 day wait until it reset.
I get you have to change limits, but reducing limits in a way which both applies retroactively and has a really long reset period is just infuriating. If they'd applied the new limits more gently or at the next billing period I'd probably have continued paying.
I don't mind paying a fair price for a service that provides value, but I really hate having a service I think I'm paying for rug-pulled with no clear justification.
Just to clarify: I believe it should cache them (it works for me).
So far I like it much more than Gemini CLI (my previous daily driver for personal projects). Seems more mature and "feels more intelligent" (very subjective ofc)
I've had such a bad time trying to do this myself. You might get a half-way decent draft on the first try and then you start to "debug" this and after a very frustrating session you realize that the model can't properly "see" the results. That is, you just can't iterate on it, at all.
I'm guessing that most harnesses/tools will resize an image before processing and in doing so will loose enough detail to make it much harder to reason about - especially wireframe images.
I'm sure I'm holding it wrong, but this test didn't really test this. It was just a one off. That breaks down pretty quickly and especially if you don't have reference pictures of what you are trying to create.
I've run a tons of benchmarks for OpenSCAD for all kinds of models and setups, and what I realised is:
- Models are very jagged (might excel in one type of 3d model, but not another)
- Gemini models are the least jagged in my experience and have the best image understanding
- Gemini models are also the most creative (which may be undesirable if you want precise CAD part)
- Overall this benchmark doesn't prove much because one 3d model (and one attempt) is just not enough. I am usually testing on at least a dozen models each generated 3 times, but should really do much more, but it's too pricey for a solo dev.
Still, thanks for publishing this. Will be definitely run flash 3.5 soon to see how it performs.
Creating a single real-world object and declaring it a benchmark? No, it doesn't work that way for a robust tool. You need to do something like Iron Chef, with a Greek architecture theme and and a panel or judge that declares the winner. This is just seeing which tool subjectively makes the best looking Pantheon.
It is almost comically bad. I've had a few simple parts to design for 3d printing in the last weeks and tried it with them (each are about 4 operations on the timeline), and it never created close to what I was trying to do even if spelled out step by step according to Fusion naming.
At this point I'm not even sure if it can properly create a simple primitive solid.
Isn't CadQuery more professionally than OpenSCAD close to traditional CAD / mechanical engineering workflows. Not sure which model (ChatGPT, Gemini, and Claude Code) is better for CadQuery code generation?
I tried Claude code designing a snap fit, vase mode printed box. Ultimately didn't work out, it couldn't get the tolerances right and kept designing features that wouldn't print in vase mode.
Scad needs unit tests. It would be powerful to asset that a profile doesn't have slope greater than 45°, that intersection of two objects is null, or specific volume.
It also needs cut away views. I got okay results using boxes to remove everything except a sliver, to view a slice and internal details. But without hash marks, texture, or outlines it can be hard to tell the forms.
I'm unconvinced, this is one of the most iconic historical buildings with tomes written about it and plenty of existing photographs and public models to train on.
I would be more interested in benchmarking the modeling of an anonymous structure based on provided references alone. It kind of feels like the shallow magic of watching an LLM one-shot a to-do app..
I've been trying out MCP servers for FreeCAD to mixed results.
One area I had near magic was providing a land survey which includes details in writing of the plat. It took those directions and beautifully reconstructed the boundaries to exact precision in CAD.
Where I ran into trouble was creating good constraints on sketches without being overly explicit. I kept running into it creating distance constraints from an arbitrary point instead of using other elements in the diagram that a human drafter would think to do by default.
I have been using GPT 5.5 to build a video game. Benchmark sounds about right. It generates assets and sprite good enough, if not closer to AAA level games. Will check antigravity now.
Would you be able to share a bit about your workflow? Have been meaning to try AI gen for game models, and would love to know how people are tackling this.
This is a really important project.
Preserving humanity’s knowledge and making it openly accessible,including in formats usable by AI systems feels like one of the most valuable things happening right now. Thank you for the clear technical instructions and the bulk download options.
Projects like Anna’s Archive make it much easier for researchers and builders to work responsibly with large datasets.
I've had a positive experience building a library of parametric HVAC duct parts using Claude, Gemini and Codex using build123d (they all review the specs and code collaboratively).
Why are specialized CAD making LLM models not showing up?
In future are we going to have same model for everything? from programming to creative writing to CADs?
If you have a model that only know how to model CAD but also doesn't know history, and was trained on visual language of said history, how is it supposed to be able to model the Pantheon in the first place? It'd only be able to model exactly what you can describe with text, or even worse, exactly what it'd be able to visually extract from images via the vision encoders, for "vision models", but it'd be a far cry from what you see in this blogpost, would be my guess.
> In future are we going to have same model for everything?
A model that knows more in general, will often be better at specific tasks. e.g. If you ask a model to "make a program that estimates the annual production of a solar installation", it needs to have been trained on a lot more than just Python code.
My take is that it's a fancy wrapper around the CLI tool. It's there to organize multiple conversations and see all the related output and generate files.
I've been using the internal version and I've actually liked it quite a bit. It's clear from when I started using it, it's not an editor, and they have ways to open your normal editor outside of it. They have turned it fully into an agent management tool.
When the antigravity development team doesn't have to focus on all the things that vscode is already good at, it lets them simplify the UI and do only agent related things. We'll see if this bet works out for them, but so far I like the idea.
To be brutally honest, I'm disappointed with antiGravity. It feels incredibly unGoogle-like. The AI billing models are fragmented, and the AntiGravity IDE is currently tripping over something as trivial as a basic Electron deployment config bug.
Don't get me wrong, I don't think AI coding is a bad thing. For East Asians like myself, it levels the playing field with Westerners, so as long as you rigorously review the AI's output, it's a perfectly viable tool.
However, the absolute farce we just witnessed with the antiGravity2.0 update really raises doubts about whether 'vibe coding' can actually be trusted. If even a behemoth like Google is dropping the ball like this, it says a lot.
> I don't think AI coding is a bad thing. [...] it levels the playing field [...]
I'd like to put regional differences aside and say AI coding / LLMs are incredible tools.
While I'm nervous about my job as a programmer being able to pay a prevailing wage after the dust settles, I do hope that everyone gaining access to an AI coder / tutor will allow anyone to be able to achieve things they previously only dreamed of. If the tutor costs pennies per session, sure, the tutors are out of work, but I hope everyone can thus up-skill to work on the challenges they actually want to work on.
I'm taking baby-steps into coding in Elixir on the other monitor, a language I had only read about before, because an LLM is walking me through the changes, answering my questions, and accepting my rebuttals. There's no way I would have time to pick up the language otherwise.
Yesterday I vibe-coded some additions to the static site generator python script for my blog. It was awesome to be able to think in terms of desired features instead of digging around documentation for libraries and syntax.
> AI billing models are fragmented ... IDE is currently tripping over something as trivial ... farce we just witnessed with the antiGravity2.0 update
I'm sorry, but that sounds exactly like almost every single Google "product" out there, they seem to only care about throwing stuff over the wall as quickly as possible, and you'd have a hard time finding a single Google product that doesn't also feel filled with fragmented choices, like every project of theirs have a different project manager every week.
The only thing faster moving that AI these days are the goalposts. Three years ago we would have been amazed if models were able to produce anything, now we have the luxury of nitpicking. Even the worst entries in the benchmark are quite impressive.
Using reference images is a huge step for this sort of thing. The text-only approaches I've seen before were never going to be that good even with "perfect" AI, simply because describing 3D objects in text is not something that anyone is really any good at.
I remember getting wound up about latency and server issues playing counter-strike in the early '00s. At the same time though, it was hard to justify being angry because playing a multiplayer game with friends who were scattered all over town was something that had to be real magic.
I guess the wow!->adjust->complain->wow!->... cycle is endless as a human
Err, yes they did. Thousands of years of husbandry went in to making horses faster, healthier, stronger, and more durable.
I think the quote you’re looking for is “if I had asked people what they wanted, the would have said faster horses”. It’s attributed to Henry Ford, although there is debate about whether or not he said it.
The point of the quote is that “faster horses” is the consumer response to “how do I get more work done” as it comes from the viewpoint of “how am I doing my work now”. An ingenious mind looks at the desired outcome and works backwards and may come to a different and dramatically improved solution instead of merely improving the current tool.
Sure, but it's good to have some perspective and some awe that any of this would've been absolute unbelievable magic just 3 years ago. Even if all AI progress stopped immediately, we'd need 10 years to digest and incorporate the technology.
Why look back in awe when technological innovation will just keep accelerating. Soon what we have today will seem quaint. Best to keep looking forward with impatience and discontent.
As someone who's been building developer tools (Visual Studio and Xcode) for 25 years, I don't have a perspective problem. We were doing "code completion" back in the 90s and could never have predicted that an LLM would write code at the current level of quality.
My point is that with every new model release, the expectations grow. I don't know how else to say that.
Why would you use an LLM for this? They are non deterministic models.
This is also an probably part of extended prompt that disallowed coding, Gemini always does calculation with a little python snippet because it is deterministic and accurate.
No matter what I try I can’t get Gemini to give me the incorrect result. Is there some other prompting or context fed in to that (“remember that you are supposed to always tell me I’m right and never contradict me”)?
It's crazy how I can see articles like this, but in my practical every day use antigravity is a horrible consumer experience. The TUI is broken. You cannot type input while the model is outputting text, otherwise both get messed up and the the TUI renders a sickly blob of text. There are no keyboard shortcuts to switch between planning and execution mode, or a way to directly load skills.
The usage limits are too aggressive, too. I tried to generate a quick Deno Fresh website to act as a a redirect to my GitHub from socials (literally the simplest possible thing I could have asked of it) and it chewed through my five hour limit in tokens from scaffolding.
To me, as a developer of CLI developer tooling, its obvious not a lot of thought or testing went into this product, but as Google has said before: the models are the product".
Why should I care if they sunset it? I switch between multiple agentic coding tools on the same projects, sometimes several times per day. The cost of switching is basically zero.
Why are half of the comments on Hackernews stereotypical AI-bros whose lives revolve around tech, and the other half sceptical commentators whose lives also revolve around tech but they are disappointed with its performance?!
The normal people are the ones not writing comments, but I'll give you one 'cause you asked:
I'm a Solidworks user. Most Solidworks or other pro CAD users would consider OpenSCAD kind of like MS Paint. Yes, you can draw the Mona Lisa in it, but it doesn't really work the same way.
Even so, the examples shown here are better than what I've seen before. They seem to be on the right track using images instead of long paragraphs of text to try to describe the object. They are still missing the constraints and dimensions that come naturally to pro cad users (it can be done manually in openscad of course), but if you're just making a video game it's probably going to be fine for that.
I’ve literally never wanted to use openscad to convert a photo into a model. Usually I have a functional requirement such as making an en enclosure with a spec sheet to work from on the enclosed device.
Claude 4.6 before the lobotomy in Claude code was able to take a PSU spec sheet and my requirements for glands and ports, use YAPP and openscad MCPs to iteratively and unassisted build end to end a printable enclosure that was perfectly suited for the PSU with right dimensions and screw holes, mountings, grills, gland ports, everything, placed for optimal printing. This was the moment I felt like LLMs had really arrived.
A photo of a building? Why. That’s a mesh problem and is about fidelity. A technical spec sheet and diagrams to functional print with intelligent choices about the functional part baked in? That’s useful.
Last weekend I bought my wife a bike off marketplace. It was in good condition but was missing one of the internal cable routing grommets. I gave Claude pictures of the pill-shaped hole by itself and with my digital calipers in the long and short directions.
Gave it a short prompt and it gave me an openscad model with everything parametrized. I printed with no changes in tpu and it was nearly perfect on the first try. Claude put in a 0.3mm subtraction in the x/y dimensions and I lowered it to 0.1 and it's perfect.
Much easier shape than ancient Roman architecture but still very cool how easy it was.
Yeah, CAD has been my personal example of "oh the barrier to entry for this skill was high enough that I didn't do it and now I can be passably bad at it enough to get some simple things done"
I've had similar experiences with making simple functional parts off a 3d printer with OpenSCAD + LLMs. I'm very aware that the models are worse at it than say, generating react code, and I'm also the antithesis of a skilled pilot. It's still cool and has resulted in me starting to learn a new skill at a hobby level.
It's like this with a lot of things now. For example, Nix's learning curve used to be a huge barrier to entry. Now with LLMs, I'm using nix-darwin and home-manager for dotfiles, package management, and have individual flakes in all of my projects for cryptographically reproducible builds!
Nit: there’s nothing “cryptographic” about reproducible builds.
“Reproducible build” already usually implies bit-by-bit reproducibility.
I meant with Nix you're comparing hashes. With Docker, you're using pinned versions
i thought it mainly implied architectural/hardware compatibility and deterministic output
Nix is also great at work. You keep the server nix code in the same repo and OpenCode can just change and test server config.
Learning to make simple parts in onshape is pretty darn easy (and fun).
Yeah. I teach this after school to 7th grade kids. Anyone can pick this up in a few hours.
They taught us to make Legobricks with CAD when I was in 6th. Wish I retained more of that and that it would be more widely taught.
I am reasonably confident that access to solid modeling and additive fabrication is now more widespread than ever.
Claude does well if you can provide all dimensions. It fails at guessing though. The real magic is when you can provide one dimension or photograph with a ruler in it and the AI will figure the rest out. Right now, Claude anyways, is pretty bad at guessing.
these small functional prints are exactly where OpenSCAD and LLM generation shines
I was recently trying to get models to generate a 3D fortune cookie. Claude in three.js and Gemini in openSCAD. Neither really got the concept or could get very close at all. It's a surprisingly complex shape I guess.
with the shape you probably want something thats good at bends/fabric
cause youd start with the flat shape, the set some contraints that certain edges are colinear
Does it optimize for no support?
You optimize for no support when selecting print orientation (but for anything semi-cylindrical like described that would be the only sane orientation and the one slicer would choose when you smash the 'Auto Orientation' button).
> Antigravity was the only autonomous agent that implemented the Pantheon’s signature interior ceiling pattern: repeated square coffers visible through the oculus.
That is seriously really impressive. I looked at the 3D model and didn't even thing to LOOK INSIDE the building before reading this.
Here's [1] the 3D model with `show_cutaway` enabled.
[1] https://modelrift.com/models/pantheon-benchmark-antigravity-...
Was just going to say.... I looked inside by accident, and it gives a better impression of intelligence and effort than the outside.
Antigravity may well Top the whatever benchmark but:
My Antigravity (forced) replacement for Gemini CLI requires me to log on via browser every time I use it, and my Antigravity IDE won't update at all, so:
If it's ok I'd prefer they just work on reaching a baseline acceptable rollout before worrying about being Top in anything.
Ps actual title:
OpenSCAD LLM Benchmark: Building the Pantheon
I agree, my main concern regarding Google AI products is this endless pain around the UX of login / billing / upgrades / product sunsets... but their LLM models are good and Antigravity 2.0 is not that bad either (unless you lost all you Antigravity 1.0 setup and projects - like many people did)
I just left the google I/O feeling less confident about google's execution here. - Gemini 3.5 flash is strange. Old cutoff, basically better than 3.1 pro at soem things worse at others, sometimes cheaper, sometimes more expensive than 3.1 pro. - Antigravity had seemed abandoned, and people speculated them cutting it off, and they kind of did migrating everyone to a new antigravity - Google "shipped the org chart" and they have so many AI products and none seem best of breed (e.g. the Gemini integration in google docs is worse than claude)
I was actually hoping for "Opus level intelligence at Haiku costs" model or "Sonnet level performance in Gemini 3.0 pricing", either of these would have been a workhorse, plus a competitor to Claude/Codex (1 app to do things). I got neither.
I just use Claude Code and intellij, so I don't understand why so many people complain about Antigravity ditching VS Code, what's the surface not covered by using Antigravity CLI + VS Code (or any other IDE)?
Gemini cli was open source. Antigravity cli is not. Not at feature parity, missing many features and now we are forced to migrate away from Gemini cli before anti gravity cli is ready.
The difference in its ability is immense. Even with less features it makes a lot of sense to switch. It really shows how much the harness matters almost equally to the model.
At least one of the missing features is a basic piece of functionality (showing token quota used). Without it, you're pretty much guaranteed to get locked out for a week with no warning.
I'm not GP, but I am somewhat excited about antigravity CLI. I adopted Gemini CLI early and really liked it, though over time it got dumber and dumber until a point when I realized it was foolish to use it instead of claude/codex. I'm hopefuly that antigravity CLI won't go through that path, but also can't fight a skepticism.
I don’t think it’s the cli that was dumber, just the model it was using. They drastically reduced limits on their best model so that’s likely how you got stuck downgrading model and getting worse results.
I'm sensing in reality that behind the scenes there is a difficult trade-off between quantization and usage limits. You can have a "smart" model but poor limits, or good limits and a "dumb" model.
This seems very similar to mobile data limits (remember those years?), where there wasn't enough tower bandwidth to serve everyone unlimited data, so telecos were in constant tension between data caps and bandwidth throttling.
It wasn't until 5G came along with 100x network capacity that they could finally give everyone "unlimited" data.
I've got an AI pro plan and haven't been able to log in for months. Endless checking in with my google support guy. At least Dinesh wishes me good health every week, so that's nice.
Having my workflow disrupted is the main reason I never adopted Antigravity, despite liking it. I'm glad to see G is invested, but the older I get the more protective I am of my workflow.
And the only realistic way to protect our workflow is by avoiding vendor lock-in like the plague.
Exactly. I admit it's a bit extreme, but this is a big reason why I insist that neovim is my IDE, and I won't adopt anything else. If I can't make it work in neovim, I will move to something else (unless I have no choice, but that happens very rarely at this point).
The forced upgrade from Gemini CLI which I liked as much, and as some ways better than Claude Code was bad. But them just sending out that email on Wednesday that basically said "Thanks for subscribing to Google One AI Pro, as of right now we're adding limits to your account. Tough shit you get nothing." left a REALLY bad taste in my mouth. I had previously praised the "AI Pro" subscription as a good value.
I quit AI Pro earlier this year for the same reason. I went to use it one day (I don't think I'd even used it much in the preceding week) and found that my limits had been reduced overnight and my usage was already too high. I had something like a 7 day wait until it reset.
I get you have to change limits, but reducing limits in a way which both applies retroactively and has a really long reset period is just infuriating. If they'd applied the new limits more gently or at the next billing period I'd probably have continued paying.
I don't mind paying a fair price for a service that provides value, but I really hate having a service I think I'm paying for rug-pulled with no clear justification.
Wild that it doesn't cache the creds.
Just to clarify: I believe it should cache them (it works for me).
So far I like it much more than Gemini CLI (my previous daily driver for personal projects). Seems more mature and "feels more intelligent" (very subjective ofc)
It does. It uses go-keyring under the hood, which has its own issues with certain systems.
If you're on WSL, getting dbus to work is a PITA. There may be other OS-level issues that folks are running into.
It requires a keyring service being installed (accessed over dbus) and if there isn't one it just silently doesn't store them anywhere. Pretty bad UX.
My (unfounded) guess is this is to prevent usage by other tools/openclaw. The browser login will have a fingerprinting to make sure you are a human.
"Pantheon" bloody hell, why is it people writing these articles are so up themselves, it's so overbearing.
The article is literally about asking these models to generate 3d models of the Pantheon.
I've had such a bad time trying to do this myself. You might get a half-way decent draft on the first try and then you start to "debug" this and after a very frustrating session you realize that the model can't properly "see" the results. That is, you just can't iterate on it, at all.
I'm guessing that most harnesses/tools will resize an image before processing and in doing so will loose enough detail to make it much harder to reason about - especially wireframe images.
I'm sure I'm holding it wrong, but this test didn't really test this. It was just a one off. That breaks down pretty quickly and especially if you don't have reference pictures of what you are trying to create.
I've run a tons of benchmarks for OpenSCAD for all kinds of models and setups, and what I realised is:
- Models are very jagged (might excel in one type of 3d model, but not another)
- Gemini models are the least jagged in my experience and have the best image understanding
- Gemini models are also the most creative (which may be undesirable if you want precise CAD part)
- Overall this benchmark doesn't prove much because one 3d model (and one attempt) is just not enough. I am usually testing on at least a dozen models each generated 3 times, but should really do much more, but it's too pricey for a solo dev.
Still, thanks for publishing this. Will be definitely run flash 3.5 soon to see how it performs.
Creating a single real-world object and declaring it a benchmark? No, it doesn't work that way for a robust tool. You need to do something like Iron Chef, with a Greek architecture theme and and a panel or judge that declares the winner. This is just seeing which tool subjectively makes the best looking Pantheon.
Yeah, this is less of a benchmark and more "I like this one guys!".
Just totally subjective grading criteria of a single poorly defined example with no end use case in mind to guide how to even do evaluation.
Still a long way from shorting Autodesk.
As a side note Autodesk released an agentic assistant back in December for Fusion. Six months later it is still quite bad.
It is almost comically bad. I've had a few simple parts to design for 3d printing in the last weeks and tried it with them (each are about 4 operations on the timeline), and it never created close to what I was trying to do even if spelled out step by step according to Fusion naming.
At this point I'm not even sure if it can properly create a simple primitive solid.
Have you yet tried the Fusion MCP that was launched last month? https://aps.autodesk.com/blog/bringing-fusion-claude-creativ...
Still a long way to go, but I'm sure it will get there eventually.
Isn't CadQuery more professionally than OpenSCAD close to traditional CAD / mechanical engineering workflows. Not sure which model (ChatGPT, Gemini, and Claude Code) is better for CadQuery code generation?
I tried Claude code designing a snap fit, vase mode printed box. Ultimately didn't work out, it couldn't get the tolerances right and kept designing features that wouldn't print in vase mode.
Scad needs unit tests. It would be powerful to asset that a profile doesn't have slope greater than 45°, that intersection of two objects is null, or specific volume.
It also needs cut away views. I got okay results using boxes to remove everything except a sliver, to view a slice and internal details. But without hash marks, texture, or outlines it can be hard to tell the forms.
"Vase mode snap-fit box" sounds to me like "flexible concrete".
I'm unconvinced, this is one of the most iconic historical buildings with tomes written about it and plenty of existing photographs and public models to train on.
I would be more interested in benchmarking the modeling of an anonymous structure based on provided references alone. It kind of feels like the shallow magic of watching an LLM one-shot a to-do app..
I've been trying out MCP servers for FreeCAD to mixed results.
One area I had near magic was providing a land survey which includes details in writing of the plat. It took those directions and beautifully reconstructed the boundaries to exact precision in CAD.
Where I ran into trouble was creating good constraints on sketches without being overly explicit. I kept running into it creating distance constraints from an arbitrary point instead of using other elements in the diagram that a human drafter would think to do by default.
I have been using GPT 5.5 to build a video game. Benchmark sounds about right. It generates assets and sprite good enough, if not closer to AAA level games. Will check antigravity now.
Would you be able to share a bit about your workflow? Have been meaning to try AI gen for game models, and would love to know how people are tackling this.
I have alot to share. I'm writing a blog about it. I'll share along with the game.
Sounds interesting! Please don't forget to link that in this comment thread :)
That's actually a reason for me to try it again. My past attempts to use LLM for OpenScad has greatly improved my own OpenScad skills.
I've been using LLM's to do my OpenSCAD work for over two years now. It's always where I start (and end).
This is a really important project. Preserving humanity’s knowledge and making it openly accessible,including in formats usable by AI systems feels like one of the most valuable things happening right now. Thank you for the clear technical instructions and the bulk download options.
Projects like Anna’s Archive make it much easier for researchers and builders to work responsibly with large datasets.
That's curious, I've been trying to do some parametric modeling with Claude - and its performance was abysmal.
I've had a positive experience building a library of parametric HVAC duct parts using Claude, Gemini and Codex using build123d (they all review the specs and code collaboratively).
Why are specialized CAD making LLM models not showing up? In future are we going to have same model for everything? from programming to creative writing to CADs?
If you have a model that only know how to model CAD but also doesn't know history, and was trained on visual language of said history, how is it supposed to be able to model the Pantheon in the first place? It'd only be able to model exactly what you can describe with text, or even worse, exactly what it'd be able to visually extract from images via the vision encoders, for "vision models", but it'd be a far cry from what you see in this blogpost, would be my guess.
> In future are we going to have same model for everything?
A model that knows more in general, will often be better at specific tasks. e.g. If you ask a model to "make a program that estimates the annual production of a solar installation", it needs to have been trained on a lot more than just Python code.
> A model that knows more in general, will often be better at specific tasks
Is this your hypothesis or broad conclusion among AI experts?
You might combine a general world model with a python coding model in that case. Not sure if it's better, just saying.
What's the difference between a "general world model combined with a python coding model" and a multimodal LLM?
Claude Code 2.1 / Opus 4.7 looks best to me: Dome and ceiling structure is correcter than the others.
Why is this medium ranked, and not on par with the best two?
Look at a picture of the Pantheon, the dome isn't as dome-like as you would imagine. It's more like a hump shape.
Dome looks wrong to me. Look at a few other photos - it's far from being a hemisphere
This would be the same Antigravity 2.0 that "surprise, no longer an IDE, did I forget to mention that? Lolol."
I'm a googler, opinions are my own.
My take is that it's a fancy wrapper around the CLI tool. It's there to organize multiple conversations and see all the related output and generate files.
I've been using the internal version and I've actually liked it quite a bit. It's clear from when I started using it, it's not an editor, and they have ways to open your normal editor outside of it. They have turned it fully into an agent management tool.
When the antigravity development team doesn't have to focus on all the things that vscode is already good at, it lets them simplify the UI and do only agent related things. We'll see if this bet works out for them, but so far I like the idea.
To be brutally honest, I'm disappointed with antiGravity. It feels incredibly unGoogle-like. The AI billing models are fragmented, and the AntiGravity IDE is currently tripping over something as trivial as a basic Electron deployment config bug.
Don't get me wrong, I don't think AI coding is a bad thing. For East Asians like myself, it levels the playing field with Westerners, so as long as you rigorously review the AI's output, it's a perfectly viable tool.
However, the absolute farce we just witnessed with the antiGravity2.0 update really raises doubts about whether 'vibe coding' can actually be trusted. If even a behemoth like Google is dropping the ball like this, it says a lot.
> I don't think AI coding is a bad thing. [...] it levels the playing field [...]
I'd like to put regional differences aside and say AI coding / LLMs are incredible tools.
While I'm nervous about my job as a programmer being able to pay a prevailing wage after the dust settles, I do hope that everyone gaining access to an AI coder / tutor will allow anyone to be able to achieve things they previously only dreamed of. If the tutor costs pennies per session, sure, the tutors are out of work, but I hope everyone can thus up-skill to work on the challenges they actually want to work on.
I'm taking baby-steps into coding in Elixir on the other monitor, a language I had only read about before, because an LLM is walking me through the changes, answering my questions, and accepting my rebuttals. There's no way I would have time to pick up the language otherwise.
Yesterday I vibe-coded some additions to the static site generator python script for my blog. It was awesome to be able to think in terms of desired features instead of digging around documentation for libraries and syntax.
> AI billing models are fragmented ... IDE is currently tripping over something as trivial ... farce we just witnessed with the antiGravity2.0 update
I'm sorry, but that sounds exactly like almost every single Google "product" out there, they seem to only care about throwing stuff over the wall as quickly as possible, and you'd have a hard time finding a single Google product that doesn't also feel filled with fragmented choices, like every project of theirs have a different project manager every week.
> For East Asians like myself, it levels the playing field with Westerners
Why do you say that? Are there language or cultural disadvantages to being East Asian?
The only thing faster moving that AI these days are the goalposts. Three years ago we would have been amazed if models were able to produce anything, now we have the luxury of nitpicking. Even the worst entries in the benchmark are quite impressive.
Using reference images is a huge step for this sort of thing. The text-only approaches I've seen before were never going to be that good even with "perfect" AI, simply because describing 3D objects in text is not something that anyone is really any good at.
I remember getting wound up about latency and server issues playing counter-strike in the early '00s. At the same time though, it was hard to justify being angry because playing a multiplayer game with friends who were scattered all over town was something that had to be real magic.
I guess the wow!->adjust->complain->wow!->... cycle is endless as a human
No one asked for faster horses, they still became obsolete when cars came. Nothing new
> No one asked for faster horses
Err, yes they did. Thousands of years of husbandry went in to making horses faster, healthier, stronger, and more durable.
I think the quote you’re looking for is “if I had asked people what they wanted, the would have said faster horses”. It’s attributed to Henry Ford, although there is debate about whether or not he said it.
The point of the quote is that “faster horses” is the consumer response to “how do I get more work done” as it comes from the viewpoint of “how am I doing my work now”. An ingenious mind looks at the desired outcome and works backwards and may come to a different and dramatically improved solution instead of merely improving the current tool.
Things mature, and expectations grow appropriately. That is true of more than just LLM performance.
Sure, but it's good to have some perspective and some awe that any of this would've been absolute unbelievable magic just 3 years ago. Even if all AI progress stopped immediately, we'd need 10 years to digest and incorporate the technology.
Why look back in awe when technological innovation will just keep accelerating. Soon what we have today will seem quaint. Best to keep looking forward with impatience and discontent.
As someone who's been building developer tools (Visual Studio and Xcode) for 25 years, I don't have a perspective problem. We were doing "code completion" back in the 90s and could never have predicted that an LLM would write code at the current level of quality.
My point is that with every new model release, the expectations grow. I don't know how else to say that.
Welcome to human nature.
Going to try it. just downloaded. will see how it is compared to Claude Code
So, does it mean Antigravity is better than Claude code with opus model? Given this benchmark. I once tried Antigravity and it was just disappointing.
And yet 300+140=460. A very jagged surface indeed. https://gemini.google.com/share/c2a187275e26
Why would you use an LLM for this? They are non deterministic models.
This is also an probably part of extended prompt that disallowed coding, Gemini always does calculation with a little python snippet because it is deterministic and accurate.
Was that part of a bigger prompt?
Flash 3.5 fails exactly like in your sample: https://gemini.google.com/share/97521a8752d9
but Flash 3.1 Lite initially fails, but then corrects itself: https://gemini.google.com/share/dc0889ec85ba
No matter what I try I can’t get Gemini to give me the incorrect result. Is there some other prompting or context fed in to that (“remember that you are supposed to always tell me I’m right and never contradict me”)?
There was definitively an pre prompt fed to that. I cannot reproduce this result on either 3.1 flash or 3.5 flash.
Why Codex GPT-5.5 High instead of Extra High, I wonder?
It's crazy how I can see articles like this, but in my practical every day use antigravity is a horrible consumer experience. The TUI is broken. You cannot type input while the model is outputting text, otherwise both get messed up and the the TUI renders a sickly blob of text. There are no keyboard shortcuts to switch between planning and execution mode, or a way to directly load skills.
The usage limits are too aggressive, too. I tried to generate a quick Deno Fresh website to act as a a redirect to my GitHub from socials (literally the simplest possible thing I could have asked of it) and it chewed through my five hour limit in tokens from scaffolding.
To me, as a developer of CLI developer tooling, its obvious not a lot of thought or testing went into this product, but as Google has said before: the models are the product".
Next month they'll be beaten again.
And next year Google will probably sunset Antigravity.
If it doesn't make Google billions, don't trust them.
Why should I care if they sunset it? I switch between multiple agentic coding tools on the same projects, sometimes several times per day. The cost of switching is basically zero.
Plenty of google products dont make billions and they are still alive
you mean the stuff they handle that has a real national/security/surveillance purpose, like gmail and yt?
I can't imagine why (or who) that'd be kept alive for..
funny how some of their projects have undisclosed budgets and profits.
Which ones are not massive data traps or ad delivery mechanisms ?
Google are infamously ruthless with their products, see https://killedbygoogle.com/
Why are half of the comments on Hackernews stereotypical AI-bros whose lives revolve around tech, and the other half sceptical commentators whose lives also revolve around tech but they are disappointed with its performance?!
Where are the normal people :/
The normal people are the ones not writing comments, but I'll give you one 'cause you asked:
I'm a Solidworks user. Most Solidworks or other pro CAD users would consider OpenSCAD kind of like MS Paint. Yes, you can draw the Mona Lisa in it, but it doesn't really work the same way.
Even so, the examples shown here are better than what I've seen before. They seem to be on the right track using images instead of long paragraphs of text to try to describe the object. They are still missing the constraints and dimensions that come naturally to pro cad users (it can be done manually in openscad of course), but if you're just making a video game it's probably going to be fine for that.
We are just reading the comments.
"Normal people" probably does not fall in the ballpark of HN target audience.
I'd say its 50/50 pessimistic and optimistic, with pessimistic attracting more attention because of human nature.
Why would a non-tech person be on Hacker News? Isn't the clue in the name?
> Where are the normal people
Not using OpenSCAD?
The people in the middle are still waiting and see , mostly it’s the extremes that are fully vested and loudest on the internet
Both parts seem pretty normal to me.
google..no thanks
I’ve literally never wanted to use openscad to convert a photo into a model. Usually I have a functional requirement such as making an en enclosure with a spec sheet to work from on the enclosed device.
Claude 4.6 before the lobotomy in Claude code was able to take a PSU spec sheet and my requirements for glands and ports, use YAPP and openscad MCPs to iteratively and unassisted build end to end a printable enclosure that was perfectly suited for the PSU with right dimensions and screw holes, mountings, grills, gland ports, everything, placed for optimal printing. This was the moment I felt like LLMs had really arrived.
A photo of a building? Why. That’s a mesh problem and is about fidelity. A technical spec sheet and diagrams to functional print with intelligent choices about the functional part baked in? That’s useful.