Introducing Physical-Engineering's Last Exam for agents
Published 8 November 2026
An 85-task benchmark found that all 14 frontier models scored below 50%, while only 41.3% of the leading model's deliverables passed every safety check.
Read the original on cloud.google.com →