Agents' Last Exam (ALE): Frontier Agents Pass Just 2.6% of the Hardest Economically-Valuable Tasks

arXiv·low signal

ALE (arXiv 2606.05405), built with 250+ industry experts, evaluates agents on 960 expert-authored, deterministically-scored workflows (1,490 instances) across 55 digital industries mapped to the O*NET/SOC occupational taxonomy. Frontier agents average 26% across all tiers but only 2.6% on the hardest 'last-exam' tier — and Codex with GPT-5.5, which scores 82% on Terminal-Bench, passes 0% of Last-Exam tasks. The benchmark is the most serious attempt yet to ground 'job-ready agent' claims in real professional work, and it's designed as a living, continuously-expanding pool.