Terminal Bench
Integrity rank #39 of 65 · 35 models scored · top score 84.72 · GPT-5.5
End-to-end agent capability on real terminal tasks (compile, train models, configure servers) in a real shell sandbox.
Strongest on discrimination, weakest on contamination resistance. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | GPT-5.5 | 84.72 | unverifiedT1 | 2026-04-23 |
| 2 | GPT-5.4 | 81.8 | unverified· optimizedT1 | 2026-03-05 |
| 3 | Gemini 3.1 Pro | 80.22 | unverifiedT1 | 2026-02-19 |
| 4 | Claude Opus 4.7 | 80.22 | unverifiedT1 | 2026-04-16 |
| 5 | Claude Opus 4.6 | 79.8 | unverified· optimizedT1 | 2026-02-05 |
| 6 | GPT-5.3 Codex | 78.43 | unverifiedT1 | 2026-02-05 |
| 7 | Gemini 3 Pro | 69.43 | unverifiedT1 | 2025-11-18 |
| 8 | GPT-5.2 | 64.94 | unverifiedT1 | 2025-12-11 |
| 9 | Gemini 3 Flash | 64.27 | unverifiedT1 | 2025-12-17 |
| 10 | Claude Opus 4.5 | 63.15 | unverifiedT1 | 2025-11-24 |
| 11 | Grok 4.20 | 57.3 | unverifiedT1 | 2026-02-17 |
| 12 | Claude Sonnet 4.6 | 53.43 | unverifiedT1 | 2026-02-17 |
| 13 | GLM-5 | 52.42 | unverifiedT1 | 2026-02-11 |
| 14 | GPT-5 | 49.61 | unverifiedT1 | 2025-08-07 |
| 15 | GPT-5.1 | 47.6 | unverifiedT1 | 2025-11-13 |
| 16 | Claude Sonnet 4.5 | 46.52 | unverifiedT1 | 2025-09-29 |
| 17 | MiniMax-M2.7 | 45.06 | unverifiedT1 | 2026-03-18 |
| 18 | Kimi K2.5 | 43.2 | unverifiedT1 | 2026-01-27 |
| 19 | MiniMax-M2.5 | 42.7 | unverifiedT1 | 2026-02-12 |
| 20 | DeepSeek-V3.2 | 39.55 | unverifiedT1 | 2025-12-01 |
| 21 | Claude Opus 4.1 | 38 | unverified· optimizedT1 | 2025-08-05 |
| 22 | Kimi K2 Thinking | 35.73 | unverifiedT1 | 2025-11-06 |
| 23 | Claude Haiku 4.5 | 35.51 | unverifiedT1 | 2025-10-15 |
| 24 | GPT-5 mini | 34.83 | unverifiedT1 | 2025-08-07 |
| 25 | GLM-4.7 | 33.39 | unverifiedT1 | 2025-12-22 |
| 26 | Gemini 2.5 Pro (Jun 2025) | 32.64 | unverifiedT1 | 2025-06-05 |
| 27 | Kimi K2 (Jul 2025) | 27.8 | unverified· optimizedT1 | 2025-07-12 |
| 28 | Grok 4 | 27.2 | unverified· optimizedT1 | 2025-07-09 |
| 29 | Qwen 3.6 35B-A3B | 24.63 | unverifiedT1 | 2026-04-14 |
| 30 | GPT-5 nano | 21.8 | unverifiedT1 | 2025-08-07 |
| 31 | gpt-oss-120b | 18.7 | unverified· optimizedT1 | 2025-08-05 |
| 32 | Gemini 2.5 Flash (Sep 2025) | 17.1 | unverified· optimizedT1 | 2025-09-25 |
| 33 | Gemini 2.5 Flash (Jun 2025) | 17.08 | unverifiedT1 | 2025-06-17 |
| 34 | Qwen3.5-9B | 9.21 | unverifiedT1 | 2026-02-24 |
| 35 | gpt-oss-20b | 3.4 | unverified· optimizedT1 | 2025-08-05 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 35 models. The spread between the best and worst scores is 81.32 points — a wide gap that supports real separation between models, since narrow spreads would mean close scores may not reflect a genuine capability gap.
2 cited facts
This benchmark earns an integrity score of 84 (grade B) and ranks 36th among 62 benchmarks; its weakest integrity component is contamination, which narrows how much the top scores can be trusted as a discriminator of frontier models under a disclosed harness.
5 cited facts
The benchmark is not saturated, with a top score of 84.72 and 15.28 points of headroom to the ceiling, while only one model clusters within a couple of points of the top. This means there is still genuine room to separate the field at the top, so the current ranking at the top is not just noise and small differences can still be meaningful.
5 cited facts
The harness comparability label is a consistent harness, so reported scores are directly comparable as task performance under that disclosed harness, not as deployed capability. The test set privacy label is unknown, so it cannot be said whether the test set is public or held out; if it were public, high scores would merit more contamination scrutiny than a held-out set, whereas low comparability would mean scores from different harnesses should not be compared head-to-head. Qualitatively, the contamination history states that tasks and reference solutions were publicly released via GitHub and users must rely on the task set being post-training-cutoff, while the harness notes disclose the model, version, performance, optimized flag, score source, benchmark release date, and model date.
6 cited facts
This benchmark measures end-to-end agent capability on real terminal tasks such as compiling code, training models, and configuring servers within a real shell sandbox. The tasks are authored as an English instruction combined with a test script and a reference solution, and different versions vary in size and sourcing, with some being community-contributed and harder. The design assumes that a real terminal sandbox tests end-to-end agent capability beyond isolated skills and that test-script verification provides objective pass/fail outcomes, while task authoring times inform difficulty and no separate human baseline is published. The sharpest caveat is that agent harness heterogeneity can affect scoring, so results may not be directly comparable across different agent implementations.
5 cited facts
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.