OpenAI GPT-6 Astra has designed a complete physical voice device from scratch as part of a push into mechanical and electrical engineering, the research company confirmed on 3 November 2026. Alongside the internal prototype, OpenAI expanded corporate engineering trials with Mazda and motorsport team Chip Ganassi Racing to test computer-aided design and aerodynamic simulations.
Key points
- Two OpenAI staff members used GPT-6 Astra to design a voice device for the Codex agent in under a week without manually editing design outputs.
- A contract manufacturer produced prototypes in under a month, though parts required several redesign passes to meet real-world factory tolerances.
- Mazda is testing the model on vehicle aerodynamics and crash structures, while Chip Ganassi Racing converts 3D scans of race components into editable computer models.
- OpenAI hardware teams have increased their internal model usage 23-fold since February, using a benchmark called BenchCAD 2.0 to track mechanical accuracy.
OpenAI revealed that two staff members built the physical prototype as a side project to test the model’s multimodal tools and computer-use capabilities. Rather than asking engineers to write code or draw schematics by hand, the team prompted GPT-6 Astra to specify requirements, shape the exterior enclosure, draft internal support ribs, select electrical components and route circuit boards.
The test device provides a dedicated hardware terminal for users to speak with Codex, the software development assistant. Astra drafted the user interface and created the underlying firmware to link the physical buttons, audio components and microphones directly to OpenAI’s network.
The experiment built on an earlier Astra demonstration where the system placed components and routed traces inside KiCad, an open-source electronics design suite, converting circuit schematics into printed circuit boards without human intervention.
The gap between software and manufacturing
While the design phase took less than seven days, turning the files into working prototypes exposed the limits of current machine intelligence in industrial settings. A contract manufacturer built a small batch of working devices in less than a month, but getting the hardware onto the assembly line required several manual review cycles.
Astra struggled to account for factory-floor tolerances, production constraints and assembly quirks. Those errors highlighted the divide between producing an unbroken digital 3D model and delivering an assembly that a fabrication plant can build reliably at scale.
OpenAI said it is using the production issues to guide training for future models. The research group intends to develop software that understands fabrication equipment limits, physical assembly steps and component tolerances alongside drafting virtual geometries.
The push into physical engineering comes as OpenAI expands its compute footprint to support complex multimodal workflows. Earlier developments, including reports that OpenAI plans purchase deals for alternative compute accelerators, mirror the rising computing costs needed to run detailed physical simulations.
Mazda tests airflow and structural safety
Beyond small consumer electronics, OpenAI is working directly with car makers to evaluate larger mechanical systems. In Japan, Mazda has paired its vehicle aerodynamics and safety teams with OpenAI’s Forward Deployed Engineering unit.
Mazda engineers use GPT-6 Astra to explore vehicle front-end styling. The model proposes alternate nose geometries, runs airflow simulations to calculate drag coefficients, and refines the shapes without human designers having to rebuild the geometry for each simulation run.
Testing dozens of variations through simulation lets engineers pinpoint the features that control aerodynamic efficiency while revealing how much freedom stylists have to alter the car’s appearance, according to Akihiro Nakata, a principal engineer at Mazda.
The automotive team also runs crash-safety simulations alongside aerodynamic modelling. The goal is to let engineers assess how altering the curvature of a bumper or bonnet impacts pedestrian impact safety, passenger protection, drag and manufacturing costs at the same time.
Reverse-engineering race cars with Chip Ganassi Racing
OpenAI has also deepened its technical work on the track. The company widened its Chip Ganassi Racing partnership, moving beyond race-day telemetry analysis into physical part fabrication.
During its initial year, the racing collaboration focused on race strategy, pit-stop performance data and telemetry gathered from on-car sensors. The updated workflow tackles missing component files, a common problem in professional motorsport where components undergo rapid trackside modification.
Chip Ganassi Racing uses Astra to convert optical 3D scans of physical racing components into fully editable computer-aided design files and 2D fabrication blueprints. Racing engineers review each generated blueprint for structural integrity and manufacturability before machine shops cut replacement parts from metal or carbon fibre.
OpenAI acknowledged that human oversight remains mandatory across the racing workflow. The racing engineers track how often the system hallucinates mating surfaces or misjudges tolerances, feeding corrections back into OpenAI’s research pipeline.
Internal adoption and BenchCAD benchmarks
Inside OpenAI, internal hardware teams working on robotics and server infrastructure have expanded their day-to-day use of Codex and ChatGPT. The company reported that token consumption across its hardware engineering groups has grown 23-fold since February 2026.
That surge in token consumption measures compute usage rather than verified hours saved, but it reflects a shift in how OpenAI develops its own data centre components and robotics platforms. Staff use the models to draft mechanical drawings, set up fluid simulations and review part documentation.
To measure technical performance across these tasks, OpenAI uses BenchCAD 2.0. The benchmark tests an artificial intelligence model on its ability to recreate complex 3D assemblies, position parts in virtual space and map electronic trace routes on multi-layer circuit boards.
In BenchCAD 2.0 tests, GPT-6 Astra running at its highest reasoning setting achieved a score of 0.594. An unreleased internal research model recorded a score of 0.875 on the same test suite. OpenAI has not detailed the architecture of the newer system, but presented the result to show that mechanical reasoning improves as models scale.
What comes next for physical AI design
As OpenAI announced, the organisation is now inviting commercial engineering firms, industrial design practices and computer-aided engineering software vendors to join testing programmes.
The company wants software vendors to integrate models directly into commercial mechanical drafting and simulation packages. Connecting language models to simulation engines allows the software to verify its own work, checking whether a bracket bends under load or an enclosure traps heat before submitting drawings to an engineer.
Industry observers continue to watch OpenAI’s internal hardware ambitions. The company is developing its own consumer devices alongside former Apple designer Jony Ive, exploring products designed specifically around voice and ambient interaction.
Whether artificial intelligence models can eliminate manual drafting depends on their ability to master manufacturing constraints. OpenAI plans to publish further technical data from its automotive trials with Mazda later in the year.
