On 20 real engineering jobs, the best AI model scored 37%; five models scored zero.
Xplore AI's new RWE-bench tested 11 AI models on 20 engineering tasks, including crankshaft design, DDR3 board routing and turbine vane CFD, each graded by a deterministic verifier with no human or AI judge. The top model, GPT-6 Astra, averaged a 37% pass rate. The hard design work is still mostly done by people.
