OpenAI says a newer internal model scores 0.875 on its BenchCAD 2.0 engineering test, against 0.594 for GPT-6 Astra

OpenAI published a post on November 3 about using its models for physical engineering work, and it includes a score for an unreleased model. On BenchCAD 2.0, which tests whether a model can rebuild 3D parts and assemblies and identify the connections on a circuit board, OpenAI says "a newer internal model" averaged 0.875, against 0.594 for GPT-6 Astra at its highest reasoning setting. We read the post on openai.com ourselves. Every number below is OpenAI's own figure.

THE NUMBERS (OPENAI'S OWN)

- BenchCAD 2.0 mean score: 0.875 for the internal model, 0.594 for GPT-6 Astra at its top reasoning setting. OpenAI's chart plots both against a relative cost proxy, and the internal model's lowest point (about 0.76) already sits above Astra's best. - Model usage across OpenAI's hardware teams, measured in tokens, has grown 23-fold since February. - Two staff designed a test device with Astra in under a week as a side project: enclosure, circuit board, firmware and the link to Codex. A contract manufacturer built a small batch of prototypes in under a month.

WHAT OPENAI SAYS DID NOT WORK

- Getting the device ready for manufacturing still took several rounds of review. OpenAI says Astra did not consistently account for the manufacturer's production limits, tolerances or assembly details.

WHO ELSE IS INVOLVED

- Chip Ganassi Racing: an expanded collaboration turning 3D scans of existing parts into editable CAD models and 2D manufacturing drawings. - Mazda: a new research collaboration with its aerodynamics and crash-safety engineers. OpenAI says GPT-6 Astra is helping model new nose shapes and simulate airflow to cut drag.

WHY IT MATTERS

BenchCAD 2.0 is OpenAI's own test, so this is not an independent result. It is still a published score for a model OpenAI has not released, and the gap over GPT-6 Astra is wide. The same day, The Information reported that Anthropic is using Claude to design its own AI chips.

WHAT WE DO NOT KNOW

- What the internal model is called, and when, or whether, it ships. OpenAI does not say, and nothing in the post ties it to the GPT-6.2 name circulating on X. - How BenchCAD 2.0 is built and scored beyond OpenAI's one-line description. We found no public version of it. - Earlier the same day, two OpenAI API model ids, rc-ada-or-vm and rc-ada-sh-vm, sat on OpenAI's model list for about five minutes. A public mirror of that list (github.com/spullara/models) shows them added at 18:41 and 18:42 UTC and removed by 18:46 UTC, our own radar saw both at 18:45 UTC, and @LuminaBench flagged them on X. What they are is unknown, and nothing links them to this model.