benchmark.darvinyi.com
← Back to Agent Evaluations
UC Berkeley RDI (Center for Responsible, Decentralized Intelligence)2026

Agents' Last Exam (ALE)

UC Berkeley RDI benchmark evaluating computer-use agents on 1,500+ expert-sourced, real professional-project tasks across 55 subdomains and 13 industries, with deterministic outcome grading and a brutal 'Last-Exam' tier where frontier agents score ~0%.

Tasks1500
Public150
Primary MetricTiered Pass Rate (Near-Term / Full-Spectrum / Last-Exam)

Overview

Agents' Last Exam (ALE) is a large-scale benchmark from UC Berkeley's RDI (Center for Responsible, Decentralized Intelligence), led by Dawn Song's team, designed to test whether 'generalist computer-use agents' are actually job-ready rather than merely benchmark-strong. Tasks are sourced from real completed professional projects with the help of 250-300+ industry experts across 100+ institutions, mapped onto the O*NET/SOC 2018 federal occupational taxonomy, and organized into 55 non-physical professional subdomains grouped into 13 industry clusters.

Unlike benchmarks that rely on subjective human or LLM-judge scoring, ALE uses deterministic, code-based evaluators that check agent outputs against verifiable reference artifacts inside full GUI+CLI computer-use environments, following a five-gate pipeline (expert sourcing → submission → peer review → implementation → QC). Tasks are split into three difficulty tiers — Near-Term, Full-Spectrum, and Last-Exam — explicitly designed to separate 'currently automatable' work from long-horizon professional work still out of reach.

Headline results from the June 2026 release show a steep cliff: the best configuration (Codex harness with GPT-5.5) reaches 38.1% on Near-Term tasks but only 22.7% on Full-Spectrum and 0.0% on the Last-Exam tier (24.0% overall); Claude Code with Claude Fable 5 scores similarly (34.3% / 20.9% / 0.0% / 22.0% overall). The paper also highlights that Codex/GPT-5.5 scores 82% on Terminal-Bench yet collapses to near-zero on ALE's hardest professional tasks, and that per-task cost varies 4–12x across similarly-performing agents (Claude Fable 5 ≈ $15.70/task vs. Composer 2.5 ≈ $1.33/task).

How It Works

Task setup, inputs, outputs, and evaluation.

SetupFull GUI + CLI computer-use sandbox reproducing a real professional working environment (software, files, tools) tied to a specific O*NET/SOC occupation.
InputA task brief distilled from a real completed professional project, with associated source materials and resources.
OutputA concrete work artifact (e.g., a filed document, model, design, or configured system state) matching the deliverable a human professional would have produced.
EvaluationDeterministic code-based checkers compare final environment/artifact state against a verified reference outcome — no subjective grading except as last resort.
MetricBinary pass/fail per task, aggregated into tier-level and overall pass rates across Near-Term, Full-Spectrum, and Last-Exam splits.

Example Tasks

Real tasks from this evaluation system.

#1

Cross-industry professional workflow task

Domain-specific professional work (one of 55 O*NET/SOC-aligned subfields)

Agent is given a real-world professional task drawn from a completed project in a specific occupation and must produce a verifiable deliverable inside a full computer-use environment.

What the Agent Receives

Task brief plus source files/resources reproduced from the original real project.

What It Must Produce

A finished work artifact matching the structure and requirements of the original professional deliverable.

How Success Is Judged

Deterministic comparison of the agent's final environment state/artifact against the verified reference outcome.

Results

Model performance on this evaluation.

#ModelScore
1
Gpt 5 5
24.0%
2
Claude Fable 5
22.0%

Key Findings

  • On the hardest 'Last-Exam' tier, frontier agent configurations (GPT-5.5, Claude Fable 5, Composer 2.5) score at or near 0% pass rate, versus 34–40% on the Near-Term tier.

  • Best overall configuration (Codex + GPT-5.5) scores 82% on Terminal-Bench but only 24% overall on ALE, and 0% on Last-Exam — a stark benchmark-to-real-work gap.

  • Per-task cost varies 4–12x across agents with similar accuracy (Claude Fable 5 ≈$15.70/task vs. Composer 2.5 ≈$1.33/task).

  • A dominant failure mode is agents declaring success before verifying their own work, producing outputs with missing files or constraint violations.

What Makes It Unique

  • Built with 250-300+ industry experts across 100+ institutions, sourcing tasks from real completed professional projects rather than synthetic or crowdsourced scenarios.

  • Uses deterministic, code-based grading against verifiable reference outcomes instead of subjective human or LLM-judge scoring.

  • Explicit tiered structure (Near-Term / Full-Spectrum / Last-Exam) designed to track 'job-readiness' over time as a living, continuously expanding benchmark.

Controversies & Caveats

Extremely low pass rates on the hardest tier (0%) may partly reflect harness/tooling limitations rather than pure model capability, since headline numbers conflate harness and backbone model choices.

As a very new (June 2026) benchmark, independent replication and long-term leaderboard stability are still limited.

Links