Terminal-Bench

Measures: An agent works in a sandboxed terminal on hard, realistic tasks spanning software, ML, science, operations, security, hardware and media; a task passes only if all its tests pass, and the score is the percentage of task trials resolved.

Maintainer: Laude Institute and Stanford researchers with open-source contributors (tbench.ai, Harbor framework)
Introduced: 2025
Status: Active
Saturation: Headroom
Typical provenance: Mixed
In scoring set: Yes

Paper Official leaderboard

Leaderboard

Show independent results only

Every result we hold, whoever ran it. Only independent and third-party rows count towards a score.

No results to show.