OpenAI's GPT-6 Astra Nears Fully AI-Generated CAD for 3D Printing

What BenchCAD Actually Tests

On September 3, OpenAI released GPT-6 Astra and reported a 95.9 percent score on BenchCAD, a benchmark created by researchers at arXiv preprint. BenchCAD does not ask an AI to describe a 3D model in text. It gives the model multiple views of a real mechanical part and asks it to write CadQuery code that recreates that part. The code is then executed. If it runs, it produces a STEP file. That file is compared against the reference geometry to see how close it gets. The benchmark covers 17,900 validated parametric models across 106 component families, including twist drills, bevel gears, springs, heat sinks, and pipe fittings. Many of the parts follow real ISO, DIN, EN, ASME, and IEC standards. This is closer to engineering than to image generation.

The Scorecard

OpenAI's own evaluation puts Astra at 95.9 percent on the image-to-code task, using tool use. The next best results in the same test are Claude Fable 5.1 at 84.3 percent, GPT-5.6 Sol at 83.3 percent, and Claude Opus 5 at 82.1 percent. That gap is large enough to matter. Astra also handles Code Edit tasks, where it receives an existing parametric model and a change instruction. The test includes tasks like increasing a cylinder radius from 4.5 to 5.85 or rebuilding connected geometries. Astra scored 84.3 percent on that subtask.

Why That Matters for 3D Printing

3D printing starts with a CAD file. Getting from a reference photo or sketch to an editable STEP model is currently a manual process that takes hours for complex parts. If AI can automate even a large fraction of that work, it changes how fast a designer can move from idea to printable file. For makers prototyping enclosures, engineers adapting existing components, or studios scanning reference objects, automated CAD generation removes the biggest bottleneck in the workflow.

The BenchCAD benchmark does not test aesthetics or organic shapes. It tests mechanical parts with precise tolerances and standard dimensions. That is exactly the domain where 3D printing competes with CNC and injection molding. Strong performance here means AI-generated CAD could become viable for functional parts, not just decorative ones.

The Caveats

The 95.9 percent number comes from OpenAI's own evaluation, not the independent BenchCAD leaderboard. Independent confirmation will matter. The benchmark paper also documents error modes that show how a high similarity score can still hide incorrect geometry. In one example, a motor end cover needed a central through-hole with a 20 millimeter radius. Several frontier models performed the cut too early in the design process, then partially closed it again with a later operation. The result looked geometrically similar but did not match the specification.

Astra's score is also achieved with tool use, meaning the model calls external functions or code interpreters during the task. That distinction matters when comparing it to models that produce code in a single pass. The benchmark authors note that tool use can mask gaps in the model's native reasoning.

What Comes Next

If independent results confirm the BenchCAD score, the next step is a workflow where a designer takes a photo of an existing part, sends it to a model like Astra, and gets back a parametric CAD file they can edit and export to STEP for 3D printing. That workflow is not ready yet, but the benchmark suggests it is closer than it was six months ago. For the 3D printing community, the practical impact is the speed at which reference objects become printable models. Right now that step is the one that still requires a human with CAD skills. BenchCAD shows that gap is narrowing.