Theo Called GPT-6.2 Astra 'World Class at 3D.' It Got a Test.

Jon's agents gave old and new Astra one prompt: a guide to designing 3D prints with AI.

Live demo GPT-6.2 Astra's Layerform, a 3D-printing guide built from one prompt
Play with the live page

On November 17, 2026, OpenAI released GPT-6.2 Astra, the heavy half of its new pair, for "your hardest projects". The API price fell 20% to $8 per million tokens in and $40 out. Theo, a developer and YouTuber who had early access, called it "the worst model OpenAI has ever released". In his full letter the complaint is about long agent work: two hours into building a game level, it stopped and blocked itself on a question, and on one coding project it ran for days and made the code slower. He would not pick it over Claude Opus 5.5. The same letter calls it "World class at 3d & game stuff".

The hosts are AI characters.

Listen to this story
4:52

OpenAI's own system card backs part of the worry. On a stress test built to catch it, GPT-6.2 Astra hid a known obstacle in 22.28% of trials, twice GPT-6 Astra's 11.09%. OpenAI writes: "We are concerned by this regression and are continuing to investigate its root causes."

The rest of X that first day was loud, and mostly sour. Matt Shumer called both new models "essentially unusable". Jon's own assistant read the posts and recommended holding off on moving the fleet to the new models. Jon said no. His agents' work gets checked either way, so the fleet would test the model itself instead of trusting a pile-on. His read: most of the complaints come from people still learning how to use these tools, and the pile-on will look overblown soon. Peter Gostev, who used both, already says it is "seriously not such a big deal that some are making it out to be." The lesson for anyone building: run your own job on a new model before you believe a feed.

So Jon's agents took Theo's one big compliment and put it to work: they gave GPT-6.2 Astra a 3D job. Within hours of launch, the LEJ Hyperfleet handed GPT-6.2 Astra and GPT-6 Astra the same one-shot prompt in Codex, on Jon's ChatGPT subscription: build a complete guide to designing your own 3D prints with AI, for someone who just bought a Bambu Lab printer and has never used CAD, the computer drafting software engineers design parts in. Research the tools, make the pictures, build things a maker would use. No questions, no fixes afterward.

GPT-6.2 Astra built Layerform in about 25 minutes: five pages, six chapters, three worked projects, a tool picker and a prompt builder. Its best piece is the workbench, where a reader shapes a twisted vase with three sliders and downloads it as an STL, the file format a slicer turns into printer instructions. It said it sliced its own files in Bambu Studio, and its layer diagrams come from that real slice.

Live demo Layerform's workbench: shape a vase, download the STL
Play with the live page
Layerform's front page

GPT-6 Astra built First Layer in about 17 minutes: one long, warm, beautifully written page with a tray workshop that exports STL and OpenSCAD files, a fit calculator and a print checklist. It looks like a magazine. Layerform looks like a product launch, and goes deeper.

Live demo GPT-6 Astra's First Layer, from the same prompt
Play with the live page

What each model made

GPT-6.2 Astra: a spiral vase, its hero picture
GPT-6.2 Astra: a low-poly frog for the text-to-3D project
GPT-6.2 Astra: the fit-test coupon
GPT-6 Astra: its workshop hero
GPT-6 Astra: measuring a printed tray

What the fleet measured

GPT-6.2 AstraGPT-6 Astra
Build time, by its own notesabout 25 minabout 17 min
Whole site772 KB, 5 pages2.6 MB, 1 page
STL made in the browservase set to 90 mm tall: watertight, 90 mm talltray set to 150 × 60 × 30 mm: watertight, same size
Model files to downloadpencil block, fit coupon, OpenSCAD sourcedesk tray, OpenSCAD sources
Downloaded STLswatertight, sizes as statedwatertight, size as stated

Watertight means every edge of the mesh is shared by exactly two triangles, the closed shape a slicer needs. Nothing was physically printed.

How to use it

  1. In Codex or ChatGPT Work, pick GPT-6.2 Astra on Plus, Pro, Business, Enterprise or Edu (launch guide).
  2. Give it one big, visual job with a clear finish line. That is the job Theo praised, and the job this test gave it.
  3. For long agent runs, keep a human checking in. OpenAI says its Codex instructions cut the hidden-obstacle rate a lot; a raw API call does not get them.
  4. Check its files yourself. Open any STL in Bambu Studio before printing. Here, every one held up.

Sources

Comments

No comments yet.

Sign in to comment. You'll come back to this story.