GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

Berkeley's Agents' Last Exam: GPT-5.5 Leads at 24%, Every Frontier Model Scores 0% on the Hardest Tier

UC Berkeley’s Center for Responsible, Decentralized Intelligence has published Agents’ Last Exam — a benchmark designed to measure whether AI agents can complete economically meaningful professional work, not just pass tests.

The results are direct: the best frontier agents pass about 24% of tasks overall. On the hardest-tier tasks, every frontier agent tested scores 0%.

What ALE Measures

ALE covers 1,490+ long-horizon tasks across 55 non-physical occupational domains, drawn from the U.S. federal O*NET / SOC 2018 occupational taxonomy. Each task is graded by code-verified outcomes with no human judges. An agent must produce deliverables that a domain expert has already validated against objective success criteria.

The benchmark is organized in tiers. Standard tasks measure whether an agent can execute a realistic professional workflow from start to finish. The “Last Exam” tier represents the hardest professional work in each domain — the kind of task where a senior practitioner earns their salary.

Led by Dawn Song, who also built MMLU, MATH, and CyberGym, ALE launched with 1,490 tasks and is designed as a living benchmark scaling toward 5,000. It was built with 250+ domain experts across more than 100 institutions.

Leaderboard at Launch

AgentPass Rate (Overall)
GPT-5.5 (Codex harness)24.0%
Cursor Composer 2.5~22%
Claude Fable 522.0%
Claude Opus 4.8Low single digits

On the hardest “Last Exam” tier, the average full pass rate across all configurations is 2.6%. Claude Opus 4.8 and every other frontier agent tested post 0.0%.

Cost per completed task:

  • Claude Fable 5: ~$15.70
  • GPT-5.5: ~$3.80
  • Cursor Composer 2.5: ~$1.33

The SWE-Bench Gap

ALE’s design deliberately separates from SWE-bench. SWE-bench tests whether an agent can locate and fix a specific code issue in an existing repo. ALE tests whether an agent can execute a multi-week professional workflow: a complete regulatory filing, a functional analytics dashboard built from raw data, a resolved insurance claim across multiple systems.

GPT-5.5 leads ALE despite Fable 5 leading SWE-bench Verified (95.0% versus GPT-5.5’s 88.7%). Researchers attribute GPT-5.5’s edge to better performance under context fatigue across very long workflows and more reliable recovery when an intermediate step fails. Fable 5 is more accurate in the early phases of a task but loses ground in multi-step error recovery.

The cost gap is the sharper finding. GPT-5.5 delivers comparable ALE performance at roughly one-quarter the per-task cost of Fable 5. For enterprise deployments running parallel agent tasks at scale, that differential translates to millions per month.

Saturation-Resistant Design

Most agent benchmarks saturate within a year. ALE is built to avoid that. Tasks are continuously onboarded, grading is code-verified rather than judge-based, and the hardest tier is anchored to professional work that humans genuinely struggle to automate.

The 0% on the Last Exam tier is not measurement error. It is the stated intent: the work most valuable to employers is still out of reach for current frontier agents. ALE puts a number on where that ceiling sits today.