On November 3, 2026, OpenAI announced that it will begin joint research with Mazda aimed at improving vehicle aerodynamics and crash safety. The company also released results from its work with racing team Chip Ganassi Racing (CGR) on converting existing parts into CAD data, as well as from its in-house design and prototyping of an electronic device. The range of AI applications is widening: from recreating design data from physical parts, to exploring new vehicle body shapes, to designing entire devices, including circuit boards. At the same time, OpenAI's prototype did not adequately account for the tolerances and assembly conditions required for manufacturing, and it needed several rounds of revisions.
Mazda: Optimizing Body Shapes. CGR: Converting Existing Parts to CAD Data
OpenAI's research team and the team that helps customers adopt its technology will begin joint research with Mazda's aerodynamics and crash-safety engineers. The plan is to simulate airflow and the behavior of vehicle structures, and to explore ways of improving performance while also taking design and cost into account. Specifically, the teams will model new shapes for the front of the vehicle and analyze the surrounding airflow to narrow down design proposals that could reduce drag. OpenAI says it will share progress and results within a few months, but it has not announced any plans to adopt the work in production vehicles, nor any specific performance improvement figures.
The CGR project, by contrast, does not generate new shapes. It recreates design data for parts that already exist. OpenAI is working with CGR's research and development team to test a process that creates editable CAD models and 2D manufacturing drawings from 3D scan data of parts. The approach is described as especially useful for parts whose original design files no longer exist, but CGR's engineers will check the accuracy of the generated data and whether the parts can actually be manufactured.
OpenAI's collaboration with CGR initially centered on analyzing data gathered from vehicle sensors, races and pit work. The new effort extends AI support beyond that data analysis into part design and prototyping. While Mazda is looking to explore new shapes, CGR is focused on recovering design information from physical objects, showing two different ways AI can be used.
Device Design With GPT-6/Astra: Designed in Under a Week, but Manufacturing Required Revisions
As an internal experiment, OpenAI also used GPT-6/Astra to design and prototype an electronic device. The AI was asked to define the functional requirements and then design the exterior, enclosure and internal structure, along with the printed circuit board's component selection, placement and routing. The scope also covered the user interface, firmware and connection to Codex, with the goal of designing the entire device without humans directly editing the design data.
Two employees took part, carrying out the design work in under a week alongside their regular duties. A contract manufacturer then produced a small batch of prototypes in under a month. However, the under-a-week timeframe applies only to the design stage, not to the total time including manufacturing. The result was also only a small-scale prototype, not a demonstration of a commercial product or mass production.
In practice, several rounds of design review and revision were needed before the device could go into manufacturing. According to OpenAI, the model could not consistently reflect tolerances (the permitted variation in dimensions), manufacturing constraints and how parts are assembled. Even when the model produced shapes that work in CAD, they cannot necessarily be manufactured as they are. Conditions that must be met on the factory floor, such as interference between parts, machining accuracy and assembly processes, need to be checked separately.
OpenAI's robotics team and the hardware team responsible for computing infrastructure are also using Codex and ChatGPT to create and edit CAD models, produce manufacturing drawings, set up simulations and conduct design reviews. Grace Lam, an engineer who helps customers adopt the technology, said Codex has significantly sped up the work of turning ideas into finished part designs. No specific time-savings figures were given, however, and the full process through to manufacturing has not been automated.
Using AI-Generated Designs in Manufacturing Requires Verification by Engineers
The three cases OpenAI presented each apply AI to a different stage of the process. CGR focuses on recovering design data for existing parts, Mazda on exploring new body shapes, and OpenAI's internal experiment on designing and prototyping an entire device. Based on the official announcement, the table below summarizes the input, the AI's work and what needs to be verified in each case.
| Case | Input | AI's work | What needs verification |
|---|---|---|---|
| CGR part reconstruction | 3D scans of existing parts | Creating editable CAD models and 2D manufacturing drawings | Dimensional accuracy, manufacturability |
| Joint research with Mazda | Design proposals such as the front of the vehicle body | Simulating airflow and structural behavior; exploring shapes | Aerodynamic performance, crash safety, design, cost |
| OpenAI's in-house prototype | Functional requirements for the device | Designing the enclosure, circuit board, software and more | Tolerances, manufacturing constraints, assembly method |
In none of these cases can the AI-generated design data or analysis results be adopted as a final deliverable as they are. At CGR, engineers must check the accuracy and manufacturability of the reconstructed CAD data, and in the Mazda research, simulation results must be evaluated against design and cost considerations. In OpenAI's in-house prototype as well, the design was revised to address manufacturing constraints. The technology outlet RuntimeWire has also pointed out that design errors translate into costs in the manufacturing process, so making design more efficient with AI and ensuring manufacturing quality need to be considered separately.
AI Usage Up 23-Fold, but Design Capability Scores Differ From Real-World Productivity
According to OpenAI, usage of AI models by its internal hardware teams working on robotics and computing infrastructure has risen 23-fold in token terms since February 2026. The absolute number of tokens was not disclosed. The figure shows growing use of AI; it does not mean that design speed or productivity has increased 23-fold.
The company also released results from BenchCAD 2.0, which measures specific CAD design capabilities. The benchmark evaluates tasks such as reconstructing 3D parts and assemblies and identifying the connections on a circuit board. The new internal model's average score was 0.875, while GPT-6 Astra scored 0.594 using its highest reasoning setting. However, these are scores on specific design tasks and do not represent the success rate or productivity of the entire manufacturing process.
The announcement shows that OpenAI is trying to extend AI's use from data analysis and software development into mechanical design and hardware development. At the same time, usage growth and benchmark scores alone cannot show how much effect there is on labor hours, development costs or production yield. The announcement also gave no pricing or availability for any dedicated design feature for general users.
What will matter in future evaluations are practical indicators: how many times engineers had to revise AI-generated designs, how much the time needed for design changes was reduced, and how the proportion of designs that meet manufacturing conditions changed. If the joint research with Mazda and the verification at CGR advance and such concrete results are published, it should become clearer how far AI can contribute not only to speeding up design work, but to making the entire development process, including manufacturing, more efficient.