LetTokenBurn
Sign in
Verified100%Harbor nativeAgentic / Tool UseCoding

Terminal-Bench

Terminal-Bench is a benchmark for measuring agents' abilities to complete tasks using a terminal.

PaperRepoDatasetLeaderboardHomepage

How to run

ltb · terminal-bench
$ npm i -g lettokenburn-cli
$ ltb bench run terminal-bench/terminal-bench@4 --benchmark terminal-bench -a claude-code -m anthropic/claude-opus-4-8
$ ltb bench push

Harbor-native: you can also call Harbor directly.

harbor
$ harbor run -d terminal-bench/terminal-bench@4 -a claude-code -m anthropic/claude-opus-4-8

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

145 sourced entries
HarnessSourceConfidenceLink
GPT-5.584.7NexAU-AHEIndependentmediumApr 23, 2026tbench.ai
GPT-5.583.1CapyIndependentmediumApr 23, 2026tbench.ai
GPT-5.582.2Codex CLIIndependentmediumApr 23, 2026tbench.ai
GPT-5.582.0CodexIndependentmediumApr 23, 2026tbench.ai
GPT-5.481.8ForgeCodeIndependentmediumMar 12, 2026tbench.ai
Claude Opus 4.780.2WOZCODEIndependentmediumApr 16, 2026tbench.ai
Gemini 3.1 Pro80.2TongAgentsIndependentmediumFeb 19, 2026tbench.ai
Claude Opus 4.679.8ForgeCodeIndependentmediumMar 12, 2026tbench.ai
GPT-5.3-codex78.4SageAgentIndependentmediumFeb 5, 2026tbench.ai
Gemini 3.1 Pro78.4Forge CodeIndependentmediumMar 2, 2026tbench.ai
Gemini 3.1 Pro78.4ForgeCodeIndependentmediumMar 2, 2026tbench.ai
GPT-5.3-codex77.3DroidIndependentmediumFeb 5, 2026tbench.ai
Claude Opus 4.676.4Meta-HarnessIndependentmediumFeb 5, 2026tbench.ai
GPT-5.3-codex75.8CodeBrain-1.5IndependentmediumFeb 5, 2026tbench.ai
GPT-5.3-codex75.7CodeliaIndependentmediumFeb 5, 2026tbench.ai
Claude Opus 4.675.3CapyIndependentmediumFeb 5, 2026tbench.ai
GPT-5.3-codex75.1Simple CodexIndependentmediumFeb 5, 2026tbench.ai
Gemini 3.1 Pro74.8Terminus-KIRAIndependentmediumFeb 19, 2026tbench.ai
Claude Opus 4.674.7Terminus-KIRAIndependentmediumFeb 5, 2026tbench.ai
GPT-5.3-codex74.6MuxIndependentmediumFeb 5, 2026tbench.ai
Claude Opus 4.672.1MAYA-V2IndependentmediumFeb 5, 2026tbench.ai
Claude Opus 4.671.9TongAgentsIndependentmediumFeb 5, 2026tbench.ai
GPT-5.3-codex71.5spoox-o-mIndependentmediumFeb 5, 2026tbench.ai
GPT-5.3-codex70.3CodeBrain-1IndependentlowFeb 10, 2026epoch.ai
Claude Opus 4.669.9DroidIndependentmediumFeb 5, 2026tbench.ai
Gemini 3 Pro69.4AnteIndependentmediumNov 18, 2025tbench.ai
GPT-5.3-codex69.1IndusAGI Coding AgentIndependentmediumFeb 5, 2026tbench.ai
Claude Opus 4.666.9CruxIndependentmediumFeb 5, 2026tbench.ai
Claude Opus 4.666.5MuxIndependentmediumFeb 5, 2026tbench.ai
GPT-5.2 Codex66.5Deep AgentsIndependentmediumJan 14, 2026tbench.ai
GPT-5.566.1clnkrIndependentmediumApr 23, 2026tbench.ai
Gemini 3 Pro65.2SageAgentIndependentmediumNov 18, 2025tbench.ai
GPT-5.2-chat64.9DroidIndependentmediumDec 11, 2025tbench.ai
GPT-5.3-codex64.7Terminus 2IndependentmediumFeb 5, 2026tbench.ai
Gemini 3 Flash Preview64.3Junie CLIIndependentmediumDec 17, 2025tbench.ai
Claude Opus 4.5 (2025-11-01)63.1DroidIndependentmediumNov 24, 2025tbench.ai
Claude Opus 4.662.9Terminus 2IndependentmediumFeb 5, 2026tbench.ai
GPT-5.2-chat62.9Codex CLIIndependentmediumDec 11, 2025tbench.ai
Gemini 3 Pro62.2CodeBrain-1.5IndependentmediumNov 18, 2025tbench.ai
Gemini 3 Pro62.2CodeBrain-1IndependentlowFeb 5, 2026epoch.ai
Gemini 3 Pro61.8II-AgentIndependentmediumNov 18, 2025tbench.ai
GPT-5.1 Codex mini61.6hookeleIndependentmediumNov 12, 2025tbench.ai
Gemini 3 Pro61.1DroidIndependentmediumNov 18, 2025tbench.ai
GPT-5.2-chat60.7MuxIndependentmediumDec 11, 2025tbench.ai
GPT-5.1 Codex Max60.4Codex CLIIndependentmediumNov 19, 2025tbench.ai
Gemini 3.1 Pro59.4Gemini CLIIndependentmediumFeb 19, 2026tbench.ai
Claude Opus 4.5 (2025-11-01)59.1Letta CodeIndependentmediumNov 24, 2025tbench.ai
claude-opus-4-5-20251101-128K59.1Letta CodeIndependentmediumDec 17, 2025tbench.ai
Claude Opus 4.5 (2025-11-01)58.4MuxIndependentmediumNov 24, 2025tbench.ai
Claude Opus 4.658.0Claude CodeIndependentmediumFeb 5, 2026tbench.ai
Top 50 of 145 by score

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

Terminal-Bench

Terminal-Bench is wide and dynamic. Terminal-Bench measures general agent capabilities on a diverse set of difficult tasks.

Terminal-Bench will be continuously updated with new and improved tasks.

Run

Terminal-Bench includes GPU and multi-container tasks. We recommend running on Modal or Daytona.

First, install the Harbor CLI.

uv tool install "harbor[modal,daytona]"
# or pip install "harbor[modal,daytona]"

And then run:

harbor run -d terminal-bench/terminal-bench \
  --agent claude-code \
  --model anthropic/claude-fable-5 \
  --n-concurrent 100 \
  --env modal

To upload results to Harbor Hub (e.g. for leaderboard submission), include the --upload flag or run harbor upload "<job/path>".

Contribute

See CONTRIBUTING.md for how to propose, implement, and submit new tasks.

We also encourage the community to reach out when they find bugs in tasks. Building benchmarks for current model capabilities is hard — community feedback helps keep quality high.