OpenAI shows CAD model: 87.5% on BenchCAD 2.0 versus 59.4% for GPT-6 Astra

If you have ever asked a language model to route a circuit board, you know how it usually ends: the layout looks tidy, and at first power-on something smells burnt. OpenAI claims it has found a way to fix this. For now these are the company's own claims, but the numbers are interesting.

What exactly was shown

OpenAI has revealed the first details of an unannounced model specialized in engineering and hardware tasks. It was tested on BenchCAD 2.0. This benchmark evaluates whether an AI can recreate 3D parts from drawings, assemble mechanical units and route electrical circuits.

The new model's result: 0.875 points, or 87.5%. The flagship GPT-6 Astra at maximum reasoning level scored 59.4%. That is a gap of almost 28 percentage points. When comparing general-purpose models with narrow specialists, that is a very noticeable difference.

Why a short circuit on the rails is no small thing

The example from the announcement itself is especially telling. The previous model made critical errors: for instance, it merged the 3.3 V, 5 V and ground lines into a single net.

For a programmer the analogy is clear. Imagine an assistant that, in one commit, merges production and test configs and also removes permission separation. The code compiles, the review looks fine, and the problem only surfaces in production. In hardware the price is higher: such a mistake means burned chips and weeks of lost time, not a commit rollback.

That is why CAD and electronics value not fluency but strict verifiability: every net in its place, every dimension adding up. This is exactly what the new model is aimed at.

The experiment and partners

In an internal experiment, OpenAI engineers designed a physical device for voice interaction with Codex in a week. The AI calculated the enclosure and internal structure, designed the board and wrote the software. After that, manual refinement and the production of the first batch of prototypes took about a month.

This proportion says a lot. Design shrank to a week, but a human, a soldering iron and manufacturing still followed. AI speeds up the start of the cycle rather than eliminating it.

OpenAI is also testing generative CAD solutions outside the lab:

• with the racing team Chip Ganassi Racing, the AI turns 3D scans of parts into editable drawings;
• together with Mazda, the model optimizes car aerodynamics by simulating airflow.

Both tasks are quite down to earth: reconstructing a drawing from a physical part and running shape variants through simulation. This is what engineers usually do slowly and by hand.

Limitations

First, what we do not know. The model's name and release date are kept secret. API, pricing and access have not been announced. The BenchCAD 2.0 figures come from the company itself, with no independent verification. It is also unclear how well the benchmark reflects real projects with complex constraints: certification, component supply, manufacturability.

The month of manual refinement after a week of AI design should also be read soberly. It is a success, but not "press a button and get a product".

What to do about it

For hardware startup founders the takeaway is simple: watch for the release and prepare processes in which an AI draft goes through rigorous checks. Electrical rules (DRC/ERC), simulation and engineer review are not going away; they become the main filter.

For product managers and developers of adjacent tools the signal matters more: big labs are moving from text and code to the physical world. If you build CAD plugins, component libraries or schematic verification systems, this market has just gained a very large player. It is better to decide in advance where you will be stronger than a general-purpose model: in data, checks or workflow integration.