Physical-Engineering's Last Exam (hidden test)
Professional Work · 2026-10-30
Physical-Engineering's Last Exam, from scientists at Google with academic partners. Agents do real mechanical, electrical, civil and structural engineering work in a full project folder, finding the governing codes and physical limits and producing deliverables a licensed reviewer would sign off on, for up to six hours per task with no internet access. An agent earns optimization points only if its design passes every hard safety check. Scripts check quantitative outputs and Claude Opus 5.5 grades open-ended work against an expert rubric. This is the mean of three runs over the 56 hidden test tasks, of 85 in all.
Benchmark source: Ple Bench