OpenAI Used GPT-6 Astra to Design a Hardware Device in One Week
OpenAI announced progress made by GPT-6 Astra in physical engineering: two employees used Astra to design from scratch a small device that can converse with Codex, with no direct manual modifications throughout the entire process; the design was completed in less than a week, and a contract manufacturer produced a small batch of prototypes in less than a month. OpenAI said the model still cannot reliably account for factory production constraints, tolerances, and assembly details, and is still several rounds of review away from mass production, so it is publicly recruiting industry partners. On BenchCAD 2.0, GPT-6 Astra scored 0.594 at its highest reasoning setting, while an unreleased internal new model averaged 0.875. OpenAI is also expanding its partnership with Chip Ganassi Racing to test converting 3D scan data into editable CAD models and 2D production drawings, and newly announced a research collaboration with Mazda on aerodynamics and crash safety.
Why read this
SelectedAI score 78
The original article provides the complete design workflow from appearance to firmware and the BenchCAD 2.0 score, which can be used to judge the model's actual boundaries in hardware design.
Topics
View original postChatGPT and Codex are helping engineers build for the physical world. We’re sharing our progress across mechanical design, electrical engineering, and simulation—from helping our hardware teams iterate faster to testing new workflows with industry collaborators. https://t.co/b3rW9zYJGA
View original
ChatGPT and Codex are helping engineers build for the physical world. We’re sharing our progress across mechanical design, electrical engineering, and simulation—from helping our hardware teams iterate faster to testing new workflows with industry collaborators. https://t.co/b3rW9zYJGA
Body · AI translation
OpenAI Moves Into Physical Engineering: GPT‑6 Astra Designs a Piece of Hardware in One Week, with Mazda and Ganassi Racing Also Taking Part in Trials
OpenAI is expanding model capabilities from writing code to mechanical, electronic, and systems design. Internally, it has already run through the complete design process from appearance, structure, and circuit boards to firmware, but it still cannot pass the "can it be mass-produced" hurdle, so it is now publicly recruiting industry partners to improve it together. https://t.co/oRvkj75Osw
Designing a device from scratch with Astra
The goal of this test was to produce a small device that can converse with Codex, without any direct human modification throughout the entire process. The functional requirements were defined by GPT‑6. The appearance was designed by Astra, referencing OpenAI's previous collaboration with Work Louder and also drawing on the style of games and toys. For the mechanical part, Astra designed the enclosure and the internal structure for mounting components. For the electronic part, it was responsible for selecting components and designing the PCB layout and routing to ensure everything could fit inside the enclosure. Finally, it also created the interaction interface and wrote the firmware and software that connect to Codex.
Two employees worked on this as a side project and completed the design in less than a week, and a contract manufacturer produced a small batch of prototypes in less than a month. However, OpenAI also admitted frankly that it is still several rounds of review and revision away from mass production: the model could not reliably account for factory production constraints, tolerances, and assembly details. This is precisely the shortcoming that the next-generation model needs to address.
How OpenAI's internal hardware team uses it
The robotics and compute infrastructure teams use Codex and ChatGPT to model and modify CAD models, convert 3D models into 2D drawings usable for production, set up simulations and finite element analysis, and also use them for design reviews. Since February of this year, the hardware team's token usage has grown 23-fold.
Three employees each described their own usage: · Grace Lam said the speed from idea to finished part has clearly increased, and drawing generation and simulation setup have saved a great deal of manual work. · Qiong Wang uses it to sort out cross-departmental design constraints, execute operations in ECAD (electronic design software), and conduct design reviews. She believes that in the future it could become a team of hardware experts, taking a circuit board all the way from concept to Gerber files ready to be delivered to a factory for production. · Jackie Lee uses it in the fields of architectural engineering and BIM (building information modeling) to automatically handle repetitive modeling, check inconsistencies between drawings and specifications, and organize issues found during coordination into to-do items that can be handled directly, with professional staff still reviewing them in the end.
Benchmark results
OpenAI uses BenchCAD 2.0 to measure progress. This benchmark examines whether a model can reconstruct 3D parts and assemblies, and whether it can identify connection relationships on circuit boards. An unreleased new internal model averaged a score of 0.875; GPT‑6 Astra scored 0.594 at its highest reasoning level. The gap is considerable, indicating that the next-generation model has made significant progress on this type of task.
Industry collaboration
OpenAI has expanded its collaboration with the Chip Ganassi Racing team, mainly testing the conversion of 3D scan data of existing parts into editable CAD models and 2D production drawings. This step is particularly useful when the original design files are lost. The team's engineers are responsible for reviewing accuracy and manufacturability, and said they can collaborate directly with Codex in CAD, with feedback changes executed directly.
In addition, OpenAI has newly announced a research collaboration with Mazda, involving its aerodynamics and crash safety teams. Astra will model new front-end styling, simulate airflow, and then refine the most promising options to reduce drag while balancing styling and cost. Mazda engineers noted that designers often ask, "Does it have to look like this?" By simulating dozens of variants, it becomes clear which features truly affect aerodynamics and where designers can freely exercise their creativity. More results will be announced in the coming months.