What the models get done.

Resolution rate is only part of the picture. Look at the agent, effort setting, token use and cost alongside the result.

Terminal-Bench 3.0

74 tasks.
A broader test of terminal work.

The July 2026 release brought a new task set covering software, science, machine learning and other domains, with reference runs and adversarial review.

Read the 3.0 release
Terminal-Bench 4.0 · latest verified release

66 tasks.
Less noise from the environment.

Released in August 2026, 4.0 removes eight tasks, fixes task issues and recalibrates compute and time limits. The official release notes set the agent timeout to eight hours.

Read the 4.0 release
Snapshot · 23 September 2026

Performance & resources

Resolution rate (%) · higher is better

Each point is a published model and agent configuration. Select a model in the table to highlight it. Full values are in the table below.0204060$0k$2.5k$5k$7.5k$10kReported full-run cost (USD)

Published configurations

Terminal-Bench 4.0 published results, checked 23 September 2026
ModelAgent / effortResolution rateRun costTokens
Codexmax58.2%±2.8 pp$3,3001.5B
Claude Codemax57.9%±3.8 pp$6,2002.7B
Claude Codexhigh53.9%±3.2 pp$6,1006.9B
Claude Codemax44.5%±3.8 pp$7,3003.8B
Claude Codemax41.8%±3.2 pp$2,7008.7B
Codexmax37.3%±3.8 pp$2,5004.4B
Claude Codemax23.6%±3.6 pp$6,5006.4B
Codexmax21.5%±3.3 pp$1,7005.7B
Grok Buildhigh20.3%±3.1 pp$3,6004B
mini-SWE-agenthigh19.1%±3.4 pp$1,80017.2B
Codexmax17.3%±2.8 pp$30011.6B
Grok Buildhigh12.4%±2.6 pp$2,1003.4B
Claude Codemax12.4%±3.1 pp$9,60021.6B
mini-SWE-agenthigh11.2%±2.4 pp$1,30011.1B

Source: Snorkel AI’s Terminal-Bench 4.0 leaderboard. Figures are a dated snapshot, not live data or Clorvin test results. Published cost figures are rounded.

Read efficiency in context.

Cost and tokens are reported run totals, not prices per task. Agent harnesses and effort settings differ. Version 4.0 also changes tasks and resource limits, so a score change from 3.0 is not a like-for-like measure of model improvement. Use the source run details before making a purchasing or model-selection decision.

Want to inspect the task side?

Task Library contains original Clorvin tasks, with new briefs, downloadable packages and locally recorded reference checks.

Explore Task Library →

Tell us what needs doing.

Share the scope, the expected volume and your timeline.
We’ll work out where our team can help.