Benchmark Registry

Terminal-Bench

Terminal-Bench results reported for 86 models across 4 versions. Updated 2026-10-08.

Report an issue

Versions

4 versions, plus 20 variants and configurations · Metric: Terminal-Bench Accuracy

VersionReleasedModelsResults
Terminal-Bench Science 0.1
1 variant of Science 0.1
VariantConfigurationReleasedModelsResults
Terminal-Bench Science 0.1 — 6x verifier timeout6x verifier timeoutSeptember 30, 202611
August 27, 2026614
Terminal-Bench 4.0Latest
7 variants of 4.0
VariantConfigurationReleasedModelsResults
Terminal-Bench 4.0 — Claude CodeClaude Code harnessAugust 28, 2026917
Terminal-Bench 4.0 — CodexCodex harnessAugust 28, 2026711
Terminal-Bench 4.0 — Grok BuildGrok Build harnessAugust 28, 202633
Terminal-Bench 4.0 — Mercor mini-swe-agentMercor mini-swe-agent, 1000 steps, 8 hoursAugust 28, 20261919
Terminal-Bench 4.0 — Muse CodeMuse Code harnessAugust 28, 202611
Terminal-Bench 4.0 — Vals AI mini-swe-agentmini-swe-agent harnessAugust 28, 20262727
Terminal-Bench 4.0 — mini-swe-agentmini-swe-agent harnessAugust 28, 202622
August 28, 2026817
Terminal-Bench 2.1
7 variants of 2.1
VariantConfigurationReleasedModelsResults
Terminal-Bench 2.1 — Thinking Machines internal coding harnessThinking Machines internal coding harnessJuly 15, 202622
Terminal-Bench 2.1 — Claude CodeClaude Code harnessMay 6, 202622
Terminal-Bench 2.1 — Mercor Terminus 2Mercor Terminus 2, 1000 steps, 3 hoursMay 6, 20261414
Terminal-Bench 2.1 — Muse CodeMuse Code harnessMay 6, 202611
Terminal-Bench 2.1 — Terminus-2Terminus-2 harnessMay 6, 202622
Terminal-Bench 2.1 — Vals AI Terminus 2Terminus-2 harnessMay 6, 20264747
Terminal-Bench 2.1 — mini-swe-agentmini-swe-agent harnessMay 6, 202633
May 6, 20261313
Terminal-Bench 2.0
5 variants of 2.0
VariantConfigurationReleasedModelsResults
Terminal-Bench 2.0 — GitHub Copilot production harnessGitHub Copilot production harnessJune 2, 202611
Terminal-Bench 2.0 — Qwen3.7 Harbor/Terminus-2, 5hQwen3.7 Harbor/Terminus-2 harness, 5hMay 20, 202622
Terminal-Bench 2.0 — Qwen3.6 Harbor/Terminus-2Qwen3.6 Harbor/Terminus-2 harnessApril 2, 202622
Terminal-Bench 2.0 — Terminus-2Terminus-2 harnessNovember 7, 202511
Terminal-Bench 2.0 — Vals AI Terminus 2Terminus-2 harnessNovember 7, 20254143
November 7, 20252530

Recently reported results: Terminal-Bench 4.0

Latest version: Terminal-Bench 4.0

ModelScoreSource
Claude Haiku 5.5 (max)39.2%anthropic.comArchived copyTerminal-Bench AccuracyReported System cardAnthropicSelf-reportedRecord
Gemini 4 Argon (highest thinking settings)57.4%deepmind.googleTerminal-Bench AccuracyReported Evaluation methodologyGoogle DeepMindSelf-reportedRecord
Claude Sonnet 5.5 (max)70.6%anthropic.comTerminal-Bench AccuracyReported System cardAnthropicSelf-reportedRecord
Claude Sonnet 5.5 (high)43.0%anthropic.comTerminal-Bench AccuracyReported Launch postAnthropicSelf-reportedRecord
Claude Sonnet 5.5 (xhigh)61.5%anthropic.comTerminal-Bench AccuracyReported Launch postAnthropicSelf-reportedRecord