TERMINAL-BENCH · benchmark platform

Terminal-Bench

Agent benchmark for hard, realistic multi-step tasks completed inside terminal environments.

verification status

verified

Last checked Oct 3, 2026

Source record

ModalitiescodeCadencerelease-basedAPIavailableEvaluations27Benchmarks1Domains1Unresolved observed labels0VerificationverifiedChecked runtime rows27Hand-checked rows0Copied rows0Backfilled0Agent rows27Checked agent rows27Task registry rows0

Copied sources mirror another provider's public page; hand-checked rows are checked against the cited page; backfilled rows are historical inserts; seeded rows are demo fixtures. Copied rows are supporting data, not first-party measurements.

TerminalBench suite

Registryterminal-bench@2.0CategoriesDifficulty

Operational state

snapshot

Latest pull

Oct 3, 2026

json

parser

Loaded 27 Terminal-Bench 4.0 benchmark records from official published rows.

0.3.0

verify

terminal-bench verification finished with status verified.

Oct 3, 2026

verified

Benchmarks from this source

Terminal-Bench 4.0

Agentic terminal coding

Accuracy

Latest change explanation

terminal-bench matched terminal-bench-20261003T114144Z with no notable change causes detected.

API notes

GET https://www.tbench.ai/leaderboard
The app ingests official published leaderboard rows server-side and preserves agent-level provenance, reasoning effort, trial counts, and confidence intervals. Earlier version 2.0 raw submissions are available from https://huggingface.co/datasets/harborframework/terminal-bench-2-leaderboard.

Method notes

Runs agents against task-specific terminal environments with test-based verification of outcomes.
Each benchmark version has its own task set and comparison group; the current official leaderboard is version 4.0.
Useful for measuring long-horizon CLI execution, not just patch generation in isolated repos.
Scores can still vary materially with the agent scaffold, tool setup, and runtime harness.
Raw submissions and per-trial files are published in the harborframework Terminal-Bench 2 leaderboard dataset on Hugging Face.

Recent source runs

Oct 3, 2026: verified · terminal-bench-20261003T114236Z
Oct 3, 2026: n/a · terminal-bench-20261003T114144Z
Oct 3, 2026: verified · terminal-bench-20261003T052207Z

TERMINAL-BENCH · benchmark platform

Terminal-Bench

Agent benchmark for hard, realistic multi-step tasks completed inside terminal environments.

verification status

verified

Last checked Oct 3, 2026

Source record

TerminalBench suite

Registryterminal-bench@2.0CategoriesDifficulty

Operational state

snapshot

Latest pull

Oct 3, 2026

json

parser

Loaded 27 Terminal-Bench 4.0 benchmark records from official published rows.

0.3.0

verify

terminal-bench verification finished with status verified.

Oct 3, 2026

verified

Benchmarks from this source

Terminal-Bench 4.0

Agentic terminal coding

Accuracy

Latest change explanation

terminal-bench matched terminal-bench-20261003T114144Z with no notable change causes detected.

API notes

Method notes

Runs agents against task-specific terminal environments with test-based verification of outcomes.
Each benchmark version has its own task set and comparison group; the current official leaderboard is version 4.0.
Useful for measuring long-horizon CLI execution, not just patch generation in isolated repos.
Scores can still vary materially with the agent scaffold, tool setup, and runtime harness.
Raw submissions and per-trial files are published in the harborframework Terminal-Bench 2 leaderboard dataset on Hugging Face.

Recent source runs

Oct 3, 2026: verified · terminal-bench-20261003T114236Z
Oct 3, 2026: n/a · terminal-bench-20261003T114144Z
Oct 3, 2026: verified · terminal-bench-20261003T052207Z

Terminal-Bench

Source record

TerminalBench suite

Operational state

Benchmarks from this source

Latest change explanation

Loading source scorecard.

Terminal-Bench

Source record

TerminalBench suite

Operational state

Benchmarks from this source

Latest change explanation