OpenAI takes Codex from code into CAD, circuit boards and vehicle simulations

What happened?

OpenAI said on November 3 that its hardware and infrastructure teams use Codex and ChatGPT to create and revise computer-aided design models, prepare manufacturing drawings, set up simulations and review designs. The announcement spans mechanical and electrical engineering: it describes CAD, electronic design environments, cross-functional constraints and design reviews, rather than a single plug-in that turns a consumer chat account into a certified engineering workstation.

The company also published a BenchCAD 2.0 comparison. The test covers reconstruction of 3D parts and assemblies and identification of circuit-board connections. OpenAI reports a score of 0.875 for a newer unnamed internal model, compared with 0.594 for GPT-6 Astra at its highest reasoning setting. The announcement does not name the newer model, publish a model card, provide a public evaluation harness or connect the result to the alleged GPT-6.2 Astra identity circulating in community trackers.

The work is being tested with collaborators. OpenAI says Chip Ganassi Racing engineers are evaluating workflows that turn 3D scans of existing parts into editable CAD models and two-dimensional manufacturing drawings. Its Mazda collaboration pairs OpenAI teams with aerodynamics and crash-safety engineers to explore vehicle designs through airflow and structural simulations while considering styling and cost. CGR’s own account describes the broader partnership as helping engineers find useful information in racing data, which is context for the collaboration, not independent validation of the new CAD benchmark.

Source: OpenAI ↗ · Source: OpenAI ↗ · Source: Chip Ganassi Racing ↗ · Source: OpenAI ↗

The benchmark is an internal signal, not a public leaderboard

BenchCAD 2.0 is useful evidence that OpenAI is measuring engineering tasks more specifically than generic coding benchmarks. It is not enough to infer that the model can take responsibility for a production design. The published score is an OpenAI result, and the post gives only a compact description of the task. Readers cannot inspect the split, scoring harness, failure distribution or an independent rerun from the public announcement. The 0.875 number should be read as a vendor-reported comparison between two named conditions, not as a universal engineering grade.

That distinction also blocks a tempting but unsupported identification. OpenAI calls the stronger system an internal model. It does not call it GPT-6.2 Astra, assign it a public endpoint or give it a release date. A tracker sighting can establish that a label appeared somewhere; it cannot authenticate the weights behind a benchmark result. Keeping the unnamed result separate from the leak story is important because otherwise a concrete engineering number would lend false credibility to an unverified model identity.

Source: OpenAI ↗ · Source: Unite.AI ↗ · Source: OpenAI ↗

Why does it matter?

The practical change is not that every developer can now ask ChatGPT for a production-ready part. It is that a coding agent is being placed inside a workflow where geometry, electronics, simulation and manufacturing constraints meet. That creates a higher-value target for automation, but also a more demanding review problem. A plausible drawing can still violate tolerances, assembly access, thermal limits, material choices or the rules of a downstream manufacturing system. OpenAI’s own device experiment says manufacturing required multiple review rounds and that the model did not consistently account for production constraints, tolerances and assembly details. The most important evidence is therefore the combination of capability and the remaining human gate.

For engineering teams, the sensible trial is narrow and auditable: choose a task with a known reference design, record the exact model and tool configuration, preserve generated CAD and drawings, and have a qualified engineer check fit, safety and manufacturability. Measure revision time and escaped defects, not only the first output. The CGR and Mazda examples are collaboration workflows with domain engineers in the loop, not proof that an unassisted agent can certify a vehicle or board. The result may still be valuable if it moves routine reconstruction and simulation setup out of the critical path while leaving sign-off with people who understand the physical system.

The story also shows why “model” needs a provenance trail. The unnamed 0.875 system, GPT-6 Astra’s 0.594 comparison, and community references to GPT-6.2 Astra are different claims with different evidence. Combining them would make the benchmark appear more reproducible and the product appear more available than the sources support. The defensible conclusion is narrower: OpenAI is demonstrating a serious internal engineering workflow and selected partner tests, while public evidence stops short of a general CAD product or a confirmed GPT-6.2 release.

Source: OpenAI ↗ · Source: OpenAI ↗ · Source: Chip Ganassi Racing ↗ · Source: Unite.AI ↗ · Source: OpenAI ↗