The grounding number: BenchCAD 95.9%
GPT-6: Astra scores 95.9% on BenchCAD, OpenAI's CAD-reasoning evaluation — one of its strongest published domain scores, alongside FrontierMath and ExploitBench.
Example prompt
Review the attached design specification against the referenced drawing. List every dimension, tolerance, or material callout that appears inconsistent between the two documents, and cite the exact section for each.Who should look elsewhere
GPT-6: Astra does not generate or edit native CAD files; it reasons about specifications, drawings-as-text, and design constraints in natural language and image input. If you need direct CAD-file manipulation, pair it with your CAD tool's own scripting layer rather than expecting native file output.
Ready to test the workflow?
Recommended access
Use pay-as-you-go API credits for a controlled evaluation on your own workload and cost profile before committing production routing.