The whole dayTue 3 Nov 2026
AI Compass
ArchiveAbout

OpenAI details physical hardware engineering with GPT-6 Astra and BenchCAD 2.0

Two OpenAI engineers designed a working handheld Codex voice prototype with GPT-6 Astra in under a week, while an internal model reached 0.875 on BenchCAD 2.0.

Isometric CAD wireframe drawing of a handheld hardware enclosure next to OpenAI’s physical engineering title card.
Image: OpenAIOpenAI’s mechanical CAD and hardware engineering workflow overview.

OpenAI shared results on 3 November 2026 from applying ChatGPT, Codex and GPT-6 Astra to mechanical design, circuit boards, firmware and engineering simulations. Internally, OpenAI’s robotics and compute infrastructure teams use the models to build and revise computer-aided design (CAD) files, draft manufacturing drawings and configure simulations, pushing hardware-team token consumption up 23 times since February 2026.

To test end-to-end multimodal and computer-use workflows without direct human edits, two OpenAI team members used GPT-6 Astra as a side project to specify requirements and design a handheld voice device for talking to Codex, inspired by OpenAI’s Work Louder collaboration. In under a week, Astra designed the external aesthetics, mechanical enclosure, internal mounting ribs, component selection, printed circuit board layout, firmware and user interface. A contract manufacturer produced working prototypes in under a month, though preparing the files for factory production still required human review across assembly details, tolerances and manufacturing limits.

ISO 129-1 mechanical engineering detail drawing for the CODEXPET seat, rim and window enclosure in black PLA generated with GPT-6 Astra.
Mechanical enclosure detail drawing generated by GPT-6 Astra for OpenAI’s handheld Codex voice prototype.Image: OpenAI

Benchmarks and industry collaborations

To track spatial and electrical reasoning, OpenAI introduced BenchCAD 2.0, an evaluation covering 3D part and assembly reconstruction alongside circuit-board connection tracing. A newer unnamed internal OpenAI model scored 0.875 on average, ahead of 0.594 for GPT-6 Astra at its highest reasoning setting.

Externally, OpenAI expanded its collaboration with Chip Ganassi Racing to convert 3D scans of legacy race-car parts without original CAD files into editable 3D models and 2D shop drawings. OpenAI also announced a research partnership with Mazda Motor Corporation, pairing forward-deployed engineers with Mazda’s aerodynamics and crash-safety teams to model nose geometries and simulate vehicle airflow to cut drag. OpenAI opened a collaboration intake form for external engineering teams and CAD software vendors, with no standalone pricing changes.

The rest of the day

Read the whole day

Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X