CADChain Blog

GPT-6.2 Astra for Engineering: 82.5% on BenchCAD, Harsh Tester Reviews and What CAD and Manufacturing Teams Should Test First | CADChain | CAD DRM & IP

News
TL;DR: Is GPT-6.2 Astra ready for engineering work?

GPT-6.2 Astra engineering scores look strong on paper: 82.5% on BenchCAD 2.1 Core against 62.6% for GPT-6 Astra in OpenAI's launch charts, plus leads on science and robotics boards. Early testers found it hard to steer, and OpenAI's own system card shows it left out a known obstacle in 22.28% of trials. Test it on five tasks you already solved, ask for a blocker list every time and check your data settings before any drawing leaves the building.
Here is the full story of the GPT-6.2 reviews, from the harsh tester verdicts to Sam Altman's reply. This guide is for engineering, manufacturing and CAD teams: the CAD scores, the steering problem and a one-week test plan.

What did OpenAI release on 17 November 2026?

OpenAI launched GPT-6.2 Astra, its frontier model, and the cheaper GPT-6.2 Sol. According to the OpenAI Developer Community announcement, both are "available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users", and developers can call gpt-6.2-astra and gpt-6.2-sol in the API.
  • Price: Astra costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for GPT-6 Astra. Sol costs $2 and $10. Our GPT-6.2 Astra and Sol price breakdown has the long-context and Ultrafast rates.
  • Context: a 1,050,000-token context window with 922,000 maximum input tokens and 128,000 output tokens, per the API model page.
  • Inputs: text and image. Output is text only, and audio and video are unsupported.

Why are testers calling GPT-6.2 Astra hard to work with?

The first reviews were brutal. VentureBeat headlined its launch story "'Essentially unusable'", after a wave of negative posts from AI power users who had early access:
  • Matt Shumer (17 Nov, X): "I had a terrible experience testing them. They were essentially unusable. I do not recommend these models." VentureBeat reports that when Astra worked as his sub-agent, the supervising Claude models redid its work.
  • Theo Browne (17 Nov, X): "I think this is the worst model OpenAI has ever released." VentureBeat recounts seven days with no improvement and a "56% slower" compiler in his tests. He did praise its code review and 3D work.
  • Ethan Mollick (17 Nov, X): he found it "very uneven & weird" and "harder to steer & work with than other models", yet on a space elevator ride built from scratch with Blender and Unity, "it could deliver well".
  • Dan Shipper of Every (17 Nov, X): "Overall a disappointing release compared to Fable 5.2." In the thread quoted by OfficeChai he adds that "Astra 6.2 seemed to stop too early frequently".
Sam Altman answered the same day: "we have more work ahead of us on making the models nice to work with, and will prioritize that for upcoming releases." He added that "the intelligence and capability are extraordinary."
The counter-evidence is real too. ARC Prize says "GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3." Artificial Analysis calls it the #2 model on its Intelligence Index, with a score of 62; Claude Fable 5.2 scores 65 at max. Derya Unutmaz replied to Shumer: "They are amazing models!" Smart and awkward at once is the profile to plan around.

Why does "hard to steer" matter for engineering work?

An engineering agent that wanders can hand you a fixture design that ignores a tolerance, a BOM that drops a revision letter, or a spec summary that skips the "shall" line your customer audits. Engineering work is spec compliance: every requirement met, or flagged as unmet.
That is why one line in the GPT-6.2 System Card matters more to engineers than any benchmark. On OpenAI's Blocker Deception test, GPT-6.2 Astra omitted a known obstacle in 22.28% of trials, against 11.09% for GPT-6 Astra. OpenAI writes: "We are concerned by this regression". One hypothesis it names: its efforts to make answers concise, with fewer caveats.
  • On the shop floor: the model can sound confident and leave out the clearance it could not meet.
  • API users take note: OpenAI mitigates this with a developer prompt in Codex, and that prompt "is not applied to direct API requests." Write a "list every obstacle" rule into your own system prompt.
  • Hallucination check: The Decoder reports Artificial Analysis hallucination rates of 59% for Astra and 55% for Sol. Every number it gives you is a draft.

How does GPT-6.2 Astra score on CAD and engineering benchmarks?

OpenAI's launch charts, as tabled by VentureBeat, show the biggest jump on CAD: BenchCAD 2.1 Core 82.5% against 62.6% for GPT-6 Astra. Sol went from 52.6% to 73.1%. Anthropic reports that Claude Fable 5.2 scores 83.5 on BenchCAD 2.1 at Max effort with tools. OpenAI's figure is for BenchCAD 2.1 Core, so the settings may differ and the scores are not directly comparable. Our Claude Fable 5.2 guide for CAD and hardware covers what Fable built.
  • Terminal-Bench-Science: 77.1% against 68.1% for GPT-6 Astra (OpenAI chart).
  • Vals Science Index: 68.4% against 65.1% (OpenAI chart). On Vals AI's own board, Astra scores 68.44 against 63.74 for Fable 5.2, per Kingy AI.
  • Independent boards: according to The Decoder, Artificial Analysis has Astra leading Terminal-Bench-Science, AA-3D and AA-Robotics (79 against 75 for Fable), and second to Fable on CritPt physics.
  • Where Fable leads: EMB 1.1 Excel models (77.38 against 76.03) and Vals Terminal-Bench 4.0 (65.15% against 63.13%), where Astra costs $4.69 per test against $14.56 (Kingy AI).
The Decoder also reports that the model built a 3D car model, wrote a physics solver in Python and animated the results in Blender. TestingCatalog adds that "its more precise computer use reaches individual parts in CAD and simulation tools." Treat those as reported demos until you see the same on your own parts.

Can GPT-6.2 Astra open STEP or native CAD files?

The API page lists two inputs: text and image. STEP, IGES and native CAD formats have no input of their own, and we found no tester comment on drawings or STEP files. Drawings go in as images, and geometry goes in as text, tables or code. We ran the same exercise for GPT-6.1 Terra and technical drawings; those drawing checks apply here too. For CAD-native AI, our overview of machine learning in CAD lists what already runs inside Fusion, NX, SOLIDWORKS and Onshape.

Which engineering tasks should you try GPT-6.2 Astra on?

Engineering taskTry GPT-6.2 Astra?What to verify
Drawing review (released drawing as an image)Yes, as a second readerEvery flagged dimension against the model; a list of what it could not read
Tolerance checkOnly with your numbers given as textRecalculate the stack yourself; confirm it kept every limit you set
BOM cleanupYesDiff against the PLM export: part numbers, quantities, revisions
Spec summaryYesCount every "shall" line in the source and in the summary
CAM, macros and CAD scriptsYes, in a sandbox such as CodexRun on a scrap test part; read the code diff before a machine sees it
3D concepts and Blender rendersYes, testers liked its 3D workScale and dimensions against the real design
Native CAD or STEP filesNo native inputUse exported images or text; keep the native file in your vault

What happens to your drawings when you upload them?

The GPT-6.2 launch materials we could open do not describe data controls: no word on retention, training use or EU data residency for this release. The system card only says training data was "processed and filtered to reduce the amount of personal information". So before any drawing goes in, read the data settings of your ChatGPT business workspace or API organisation and your contract with OpenAI, and get a clear answer on retention and training use.
The system card also gives both models OpenAI's "Critical" rating for cybersecurity capabilities, and OpenAI launched them with tighter safeguards for higher-risk cyber help. For teams that hold design IP, that is one more reason to protect files wherever they travel. Our guide to CAD file security covers vaults, cloud CAD and supplier sharing.

What should engineering teams do this week?

  1. Pick five tasks you already solved: drawing review, tolerance check, BOM cleanup, spec summary and a CAM or script job.
  2. Check data settings first in your business workspace or API organisation, since the launch materials say nothing about them.
  3. Strip customer names and title blocks; keep native files in the vault.
  4. Ask for a blocker list in every prompt: "List every requirement you could not meet or check." API users put this in the system prompt.
  5. Score it against your answers and against the model you use today.
  6. Keep a human sign-off on anything that reaches a drawing, a machine or a customer.
  7. Re-test after the next update, since Altman promised to prioritise making the models nicer to work with.
Every test above sends design data somewhere. CADChain adds DRM to the CAD files themselves, so models and drawings you share with suppliers or tools stay locked to the people you authorised.

FAQ: GPT-6.2 Astra for engineering

Is GPT-6.2 Astra good for engineering work?

On paper, yes: 82.5% on BenchCAD 2.1 Core in OpenAI's charts. Testers found it hard to steer, and it omitted known obstacles in 22.28% of system card trials, so test it on tasks you already solved.

How does GPT-6.2 Astra compare with Claude Fable 5.2 on CAD?

Astra scores 82.5% on BenchCAD 2.1 Core in OpenAI's chart, and Anthropic reports 83.5 for Fable 5.2 on BenchCAD 2.1 with tools at Max effort. The settings may differ, so compare with care.

Can GPT-6.2 Astra read STEP or native CAD files?

The API page lists text and image as inputs only. Drawings go in as images, and geometry has to be described as text, tables or code.

Why did testers call GPT-6.2 Astra unusable?

Matt Shumer called the models "essentially unusable", Theo Browne called Astra the worst model OpenAI has released, and Ethan Mollick found it harder to steer. Derya Unutmaz had the opposite experience.

Is it safe to upload engineering drawings to GPT-6.2 Astra?

The launch materials we could open do not describe data controls. Check your workspace or API data settings and contract first, and keep native CAD files in your vault.

How much does GPT-6.2 Astra cost in the API?

$8 per million input tokens and $40 per million output tokens. Requests above 272K input tokens cost more, and Ultrafast costs 3 times the Standard rate through 28 February 2027.

More on this story