Benchmark explainer
What is Terminal Bench?
End-to-end agent capability on real terminal tasks (compile, train models, configure servers) in a real shell sandbox.
Terminal Bench as reported in Epoch AI's Capabilities Index CSV.
How it's scored
- Metric
- normalized performance in Epoch CSV, rendered as accuracy (%)
- Score ceiling
- 100
- Construction
- Tasks authored as English instruction + test script + reference solution; v2.1 has 89 tasks; v3.0 (formerly Frontier-Bench) is community-contributed and harder; v4.0 has 66 tasks.
- Human baseline
- Not separately published — task authoring times inform difficulty (real-workflow inspiration).
How trustworthy
Discrimination
100/100Does it still separate models?
Saturation headroom
97.6/100How far from ceiling / clustered at the top?
Contamination resistance
0/100Public vs held-out; training-leak risk.
Harness comparability
100/100Is it apples-to-apples, or mixed / vendor-optimized?
Freshness
91.7/100How old is the benchmark?
Contamination history: Tasks and reference solutions publicly released via GitHub; users must rely on v4.0 task set being post-training-cutoff.
Limitations: Compute and time allowances are tuned per-version; agent harness heterogeneity (Claude Code vs OpenHands vs Codex) affects scoring; held-out tasks only if not on training web.
Who leads Terminal Bench
| Model | Score | Evidence |
|---|---|---|
| GPT-5.5 | 84.72 | unverified |
| GPT-5.4 | 81.8 | unverified |
| Gemini 3.1 Pro | 80.22 | unverified |
| Claude Opus 4.7 | 80.22 | unverified |
| Claude Opus 4.6 | 79.8 | unverified |
Compare the top Terminal Bench scorers
Frequently asked questions
Terminal Bench: Terminal Bench as reported in Epoch AI's Capabilities Index CSV.
GPT-5.5 leads Terminal Bench at 84.72. The full leaderboard above lists every recorded measurement, not just the headline number.
35 models have recorded scores on Terminal Bench, spanning a score spread of 81.32.
tensor.news grades Terminal Bench B for integrity (score 83/100), ranking #39 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — Terminal Bench still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on Terminal Bench are reasonably apples-to-apples.
Every Terminal Bench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites www.tbench.ai.