Leaderboard

Terminal-Bench Hardofficial site ↗

22 models scored · metric: success rate. Prices are the cheapest listed across all serving providers. Cost/run and Pts/$ are priced for the agentic token profile — tool loops re-send a growing transcript and think between calls.

A benchmark is priced for the workload it exercises, not a fixed input:output blend: an agent loop and a single hard question bill very differently on the same rate card. Profile: 12k in · 2k out · 3k reasoning · 70% cache hit · 32k context.

Scores are facts as publicly reported by labs (see report links below and on each model page) and aggregated via models.dev. Terminal-Bench Hard is maintained by its own project — Model Pulse is not affiliated with or endorsed by it.