Breaking

New ALE benchmark for real-world agent work debuts; Claude Opus 4.8 scores 0% on hardest set

June 10, 2026 at 06:21 EDT

AI Agents
Research & Papers
Foundation Models

A research team led by UC Berkeley RDI released a new benchmark called Agents' Last Exam (ALE) around June 9, 2026, designed to measure how well AI agents perform real-world professional work. More than 300 domain experts from over 100 institutions contributed across 55 industry subfields, yet Claude Opus 4.8 scored a 0.0% pass rate on the hardest subset.

June 9, 2026 · Agents' Last Exam (ALE)

On a Benchmark of Real Jobs, Frontier AI Agents Pass Just 2.6%

ALE tests whether AI agents can finish long-horizon professional projects end-to-end in a real computer environment. On its hardest tier, the best agents averaged a 2.6% full pass rate — and Claude Opus 4.8 scored a flat 0.0%.

Hardest tier — full pass rate

100%

Human
professional

2.6%

Top agents
(average)

0.0%

Claude
Opus 4.8

Columns drawn to scale — current agents barely clear zero.

~1,500

real-world tasks
(target 5,000)

industry subfields
(13 clusters)

300+

expert authors
(100+ institutions)

49 hrs

to process one
hardest task (~3 days)

How each task is run

Real environment

Full GUI + CLI access on an actual computer

→

Long-horizon project

Real work that takes humans days to weeks, start to finish

→

Code-graded deliverable

Objective, reproducible evaluators — no human judgment

Why it matters

A serious step toward measuring economically meaningful long-horizon work — with objective scoring spanning GUI, CLI, coding, and tool calls, built on official occupational classifications.

The reality check

Near-zero pass rates show how far agents remain from doing real jobs. Runtimes of up to three days per task raise added concerns about practicality.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.