Agents' Last Exam (ALE)
ALE is a UC Berkeley (RDI) and Snorkel AI benchmark from June 2026 that evaluates agents on long-horizon, economically valuable professional tasks across 55 subfields, built with 300+ industry experts. Its public release, ALE-V1, includes 147 reference tasks from a 1,500+ task corpus (arXiv:2606.05405), with a launch pass rate of ~2.6% and a July 2026 leaderboard top score of 53.6/100 by Codex + GPT-5.6; it is a living benchmark with task rotation every 6 months.