CADBench pilot benchmark evaluates 10 AI CAD agents on 28 mechanical design tasks with layered scoring across geometry, engineering, manufacturability, and cognition.

CAD-Bench Lab released a research-grade benchmark (May 2026) running a 28-task pilot subset from a 343-task suite across 10 agents at 5 seeds each, with scoring reported as bootstrapped 95% CIs, worst-case p5, and 2PL IRT ability θ calibrated against task difficulty.

Scoring uses four layers—geometry, engineering, manufacturability, and cognition—with three use-case views that re-weight the layers dynamically; results include a (capability, $/task) Pareto frontier showing which agents are not dominated on both capability and cost axes.

The pilot evaluated 20 categories across the 4 layers, with an aggregate of 431.7 min wall-clock compute; 4 of the 10 agents were human baselines for comparison.