OpenAI evaluates specialized CAD and electronics engineering model OpenAI

OpenAI revealed preliminary evaluations of an unannounced model specialized for computer-aided design (CAD) and hardware engineering. The model achieved a 0.875 score on the BenchCAD 2.0 evaluation suite.

On BenchCAD 2.0, which tests 3D part reconstruction from blueprints, mechanical assemblies, and printed circuit board (PCB) traces, the model outperformed flagship GPT-6 Astra at max reasoning (0.594). In early tests, prior models produced critical electrical bugs, such as shorting 3.3V, 5V, and ground rails into a single net.

OpenAI validated the system by designing a handheld physical voice device for Codex in one week. The agent laid out the enclosure, internal mounting brackets, and four-layer PCB, alongside starter firmware. External partners, including Chip Ganassi Racing and Mazda, are piloting the tooling to convert 3D point-cloud scans into editable parametric CAD models and automate aerodynamic airflow meshes.

What it is

• BenchCAD 2.0 evaluation. Scored 0.875 across drafting, assembly, and schematic routing, compared to 0.594 for GPT-6 Astra.
• Electronics routing. Solves PCB routing without shorting disparate voltage rails, resolving a known failure mode in general reasoning models.
• Physical hardware pilot. OpenAI engineers used the model to design a physical voice appliance for Codex within one week.
• Industrial partnerships. Chip Ganassi Racing uses the model to turn physical scan data into editable CAD geometry; Mazda applies it to vehicle aerodynamics simulation.
• Unreleased preview. Official model branding and API availability dates have not yet been announced.

Sources

• OpenAI: Accelerating engineering for the physical world