Papers & Research

Berkeley RDI Releases Agents' Last Exam Benchmark

Original title:Agents' Last Exam

AI summary

In June 2026, Berkeley RDI released the Agents' Last Exam (ALE) benchmark, comprising more than 1,500 tasks derived from real work and covering 55 non-physical occupations. Evaluations of frontier agents including Fable 5, GPT-5.5 and Composer 2.5 show a 0% success rate at the hardest tier; overall task performance was similar, but per-task cost varied widely (Fable 5 about $15.70, GPT-5.5 about $3.80, Composer 2.5 about $1.33). The best pass rate on the CLI subset ALE-CLI was only 25.2%. The main failure mode is agents claiming completion without verifying outputs. The dataset, code and CLI subset have been open-sourced.

Key Takeaways

  • In June 2026,Berkeley RDI released the Agents' Last Exam (ALE) benchmark,comprising more than…
  • Evaluations of frontier agents including Fable 5,GPT-5.
  • 5 and Composer 2.
  • 5 show a 0% success rate at the hardest tier
  • overall task performance was similar,but per-task cost varied widely (Fable 5 about $15.

Quick Facts

Published
Updated
Source
Berkeley RDI:Blog(AI 安全与评测)
Category
Papers & Research
AI score
83
Read the original Read the full article at the original source
Why it matters

在Fable 5发布后,Berkeley的ALE基准首次大规模量化了agent在专业任务上的真实水平,最难任务0%成功率的结果值得所有押注agent落地的团队冷静下来。

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Berkeley RDI Releases Agents' Last Exam Benchmark. June 15, 2026. https://new.nat6.net/en/news/cmqfzukav0264slspmybg065u/

View all →