FrontierCAD (Normal)
A CAD benchmark from Normal Research's CAD Arena that scores how well AI agents recreate real parts across CAD tools; it is a benchmark, not a generative product.
What it does and how it connects
FrontierCAD is the verifier behind Normal Research's CAD Arena benchmark. An agent receives an engineering drawing licensed from the manufacturer that produced the part and rebuilds it as editable CAD through the scripting API of NX Open, SolidWorks, Onshape FeatureScript, Autodesk Fusion, or Build123d.
The FrontierCAD verifier scores geometric match to the reference part and whether the native sketches, features, and parameters support later edits. The combined 0 to 100 quality score is not a pass rate. The Fireworks index page, last updated October 20, 2026, lists 12 models tested, with Claude Opus 5.5 leading at 74.6 out of 100, 7.5 points ahead of GPT-6 Astra, which costs 80 percent less per task.
The CDFAM survey of AI to CAD benchmarks notes that CAD Arena, from Normal Research, tests AutoCAD, Siemens NX, SolidWorks, Onshape, and Fusion, and that it is not peer reviewed. It is a vendor-run benchmark, not an academic one, and the scores are published by the organization that runs the benchmark.
Alternatives
Compare side by side- PartForge evals
A standing CAD-agent evaluation of hard CAD tasks rerun against each model release, published as model comparisons, rather than a benchmark that scores agents recreating real parts.
- RealCADBench
A benchmark for intent-to-program CAD modeling from real industrial design intents, scored on FreeCAD Python output, rather than a benchmark that scores agents recreating parts across CAD tools.
- PCBench
A dataset of 164 KiCad boards with routing ground truth for measuring autorouter quality, rather than a benchmark that scores agents recreating parts.
- freerouting self-improving loop
An agent-driven experiment loop inside the freerouting repo that scores routing optimizer changes against frozen baselines, rather than a benchmark that scores agents recreating parts.
- Easy-CAD / bendcad (jnadeau207-collab)
An agent-native CAD kernel effort built for AI agents, with certified volumes and named refusals, rather than a benchmark that scores agents.
Sources
Not verified
- License: no license or terms page was found for this benchmark; the commercial kind follows the directory record.
- Pricing: none stated on the sources read.
- Real user voices (no real linked post read for this page)
- Related Error Index entries (no matching entry verified)
- No public code, dataset, or leaderboard endpoint was found on Normal Research's own site; the benchmark is described on the Fireworks index page, which links Normal Research as the source.
- The directory record calls the benchmark FrontierCAD; the Fireworks index page calls it CAD Arena and names FrontierCAD as its verifier. Both names are recorded here.
Last verified November 6, 2026. Listings are informational. The evidence score measures public, verifiable evidence and is not an endorsement or a compatibility claim.