OpenAI physical engineering work now includes an unreleased model that the company says scored 0.875 on BenchCAD 2.0, well above GPT-6 Astra's 0.594. The November 3 disclosure also introduced a research collaboration with Mazda and expanded a project with Chip Ganassi Racing.

The results point toward AI systems that operate across design software, simulation and manufacturing documentation. They do not establish that engineering can run without specialists: OpenAI says its Astra-designed prototype needed repeated review because the model missed production limits, tolerances and assembly details.

The OpenAI physical engineering update contains four concrete developments:

• An internal model reached 0.875 on BenchCAD 2.0.
• Astra designed a Codex-connected prototype in under a week.
• OpenAI began aerodynamics and crash-safety research with Mazda.
• Chip Ganassi Racing is testing scan-to-CAD workflows.

OpenAI Physical Engineering Model Reaches 0.875 on BenchCAD

OpenAI's primary report says a newer internal model averaged 0.875 on BenchCAD 2.0, compared with 0.594 for GPT-6 Astra at its highest reasoning setting. OpenAI did not name the internal model or provide a release date.

BenchCAD 2.0 evaluates whether models can reconstruct parts and assemblies from drawings or rendered views and infer wiring from images of a physical circuit board. Its Core set contains 100 held-out cases across six tasks, with each result checked against references rather than scored by another language model.

The benchmark's published methodology allows up to 30 rounds per case in an offline sandbox with Python, CadQuery and numerical libraries. That setup measures a model working through a technical tool chain, not simply answering an engineering question in prose.

The public BenchCAD site lists Astra near 0.59 on the Core evaluation, broadly matching the baseline in OpenAI's disclosure. The 0.875 figure, however, is a company-reported result for an unavailable system. It has not been independently reproduced, and buyers cannot yet test the claimed improvement in production.

Astra Prototype Exposed Manufacturing Limits

OpenAI asked GPT-6 Astra to define the requirements and design a small device for speaking to Codex. The work covered the enclosure, internal mechanical structures, component selection, printed circuit board layout, user interface, firmware and software integration.

Two team members completed the design in under a week as a side project, according to the company. A contract manufacturer then produced a small batch of prototypes in under a month. OpenAI describes the design itself as having no direct human edits.

That claim stops short of autonomous manufacturing. OpenAI says several review and iteration rounds were necessary before production because Astra did not reliably account for the manufacturer's limits, tolerances or assembly requirements. Those omissions can turn a plausible CAD model into a costly or unbuildable part.

The distinction matters because physical errors have different consequences from a flawed draft or broken software test. Engineers must verify clearances, material behavior, electrical safety, thermal constraints and fabrication capabilities before a design reaches a machine or production line.

Mazda and Chip Ganassi Racing Test Engineering Workflows

The Mazda research collaboration brings OpenAI's research and forward-deployed engineering teams together with the automaker's aerodynamics and crash-safety specialists. The partners are exploring airflow and structural simulations for vehicle designs while considering styling, performance and cost.

OpenAI says Astra in Codex is being used to model alternative front-end shapes, simulate airflow and refine candidates that could reduce drag. Mazda engineers remain responsible for evaluating whether the outputs are accurate, useful and compatible with real vehicle-development constraints.

With Chip Ganassi Racing, the focus is reverse engineering. Models are being tested on turning 3D scans of existing parts into editable CAD models and two-dimensional manufacturing drawings, a workflow that can help when original design files are missing.

Independent coverage from AI Magazine confirmed the announced collaborations and noted that OpenAI's internal hardware teams are also using Codex and ChatGPT for CAD revisions, drawings, simulations and design reviews. OpenAI says token usage by those teams has grown 23-fold since February.

Engineering Adoption Depends on Verification and Tool Access

The update expands the competitive field for coding agents into mechanical and electrical engineering. A useful system must interpret drawings, manipulate specialist software, preserve constraints across disciplines and produce artifacts that a human can audit. Performance in one benchmark captures only part of that workflow.

OpenAI has not disclosed the internal model's cost, latency, failure distribution or performance on production-scale assemblies. It also has not said when the model will be available. Those details will determine whether the benchmark improvement translates into faster engineering rather than additional review work.

The near-term opportunity is narrower and more credible: assist with repetitive CAD changes, drawing preparation, simulation setup and recovery of editable geometry from scans. Human sign-off remains essential where an error could waste material, damage equipment or compromise safety.

The next evidence should come from controlled external evaluations and documented deployments that measure time saved, correction rates and manufacturing outcomes. Until then, the prototype and partnerships demonstrate expanding scope, while the disclosed manufacturing misses define the present boundary.

Photo: OpenAI's headquarters at 1515 Third Street in San Francisco, photographed by Coolcaesar, via Wikimedia Commons. Licensed under CC BY 4.0. Cropped and resized to 16:9; the image does not depict the engineering tests.

Further Reading

• Qualcomm HBC Silicon Passes Validation Ahead of 2027 AI250 Sampling
• GDPx Benchmark Launches With Claude Opus 5.5 Leading at 12%