Terminal-Bench
Command-line agent tasks; best result per model across harnesses.
As of 25 Oct 2026, GPT-5.5 leads Terminal-Bench on BenchLeader with 84.7%, ahead of GPT-5.4 at 81.8%, across 84 model configurations with a published result.
- Published by
- Terminal-Benchdata via Epoch AI Benchmarking Hub
- Category
- Agents & tools
- Index weight
- 1.0
- Models
- 84
- Data as of
- 25 Oct 2026
- Newest result
- 29 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
Terminal-Bench gives an agent a shell in a container and a task such as building a project from source, recovering a corrupted file system, or setting up a service, then checks the end state with tests.
How it is scored
Percent of tasks whose checks pass. Epoch's board keeps the best published result per model across harnesses.
What to keep in mind
Harness and tooling matter a great deal, so best-per-model mixes different agents. Tasks are long, so a single flaky run moves the score; the harder subsets run by Artificial Analysis and Vals AI are more stable for ranking.
- 1GPT-5.584.7%
- 2GPT-5.481.8%
- 3Claude Opus 4.780.2%
- 4Gemini 3.1 Pro80.2%
- 5Claude Opus 4.679.8%
- 6GPT-5.3-Codex78.4%
- 7Gemini 3 Pro69.4%
- 8GPT-5.2-Codex66.5%
- 9GPT-5.264.9%
- 10GPT-5.2 (medium)64.9%
- 11Claude Opus 5.5 (max)64.8%
- 12Gemini 3 Flash64.3%
- 13Claude Opus 4.563.1%
- 14Claude Sonnet 5.5 (max)61.8%
- 15GPT-5.1-Codex-Mini61.6%
| # | ||||
|---|---|---|---|---|
| 1 |  GPT-5.5 | 84.7% | 60.4 | 2026-04-23 |
| 2 | GPT-5.4 | 81.8% | 58.1 | 2026-03-05 |
| 3 |  Claude Opus 4.7 | 80.2% | 63.2 | 2026-04-16 |
| 4 |  Gemini 3.1 Pro | 80.2% | 60.5 | 2026-02-19 |
| 5 | Claude Opus 4.6 | 79.8% | 62.5 | 2026-02-05 |
| 6 | GPT-5.3-Codex | 78.4% | – | 2026-02-05 |
| 7 | Gemini 3 Pro | 69.4% | 60.4 | 2025-11-18 |
| 8 | GPT-5.2-Codex | 66.5% | 57.3 | 2025-12-18 |
| 9 | GPT-5.2 | 64.9% | 58.0 | 2025-12-11 |
| 10 | GPT-5.2medium | 64.9% | 57.6 | 2025-12-11 |
| 11 | Claude Opus 5.5max | 64.8% | 72.6 | 2026-09-22 |
| 12 | Gemini 3 Flash | 64.3% | 57.0 | 2025-12-17 |
| 13 | Claude Opus 4.5 | 63.1% | 57.9 | 2025-11-24 |
| 14 | Claude Sonnet 5.5max | 61.8% | 69.4 | 2026-09-28 |
| 15 | GPT-5.1-Codex-Mini | 61.6% | – | 2025-11-12 |
| 16 | GPT-5.1-Codex-Maxmax | 60.5% | – | 2025-11-19 |
| 17 | GPT-6 Astramax | 58.2% | 70.4 | 2026-09-03 |
| 18 | GPT-6.1 Solmax | 58.2% | 69.9 | 2026-09-29 |
| 19 | Claude Fable 5.1max | 57.9% | 70.5 | 2026-09-01 |
| 20 | GPT-6 Astrahigh | 57.9% | 69.6 | 2026-09-03 |
| 21 | Claude Fable 5.1xhigh | 57.9% | 69.5 | 2026-09-01 |
| 22 | GPT-6 Astraxhigh | 57.9% | 68.9 | 2026-09-03 |
| 23 | GPT-5.1-Codex | 57.8% | – | 2025-11-12 |
| 24 |  Grok 4.20 | 57.3% | – | 2026-02-17 |
| 25 | Claude Fable 5.1high | 54.5% | 69.4 | 2026-09-01 |
| 26 | GPT-6 Astramedium | 54.2% | 67.5 | 2026-09-03 |
| 27 | Claude Opus 5xhigh | 53.9% | 67.4 | 2026-07-24 |
| 28 | Claude Fable 5.1medium | 53.9% | 67.3 | 2026-09-01 |
| 29 | Claude Sonnet 4.6 | 53.4% | 57.1 | 2026-02-17 |
| 30 |  GLM-5open ↗ | 52.4% | 54.9 | 2026-02-11 |
| 31 | Claude Opus 5max | 51.8% | 68.8 | 2026-07-24 |
| 32 | GPT-6 Astralow | 50.6% | 65.9 | 2026-09-03 |
| 33 | Claude Opus 5high | 50.3% | 67.9 | 2026-07-24 |
| 34 | GPT-5 | 49.6% | 56.2 | 2025-08-07 |
| 35 | GPT-5medium | 49.6% | 57.5 | 2025-08-07 |
| 36 | GPT-6 Solmax | 49.4% | 66.4 | 2026-09-22 |
| 37 | GPT-5.1medium | 47.6% | 52.3 | 2025-11-13 |
| 38 | GPT-5.1 | 47.6% | 56.8 | 2025-11-13 |
| 39 | Claude Sonnet 4.5 | 46.5% | 53.7 | 2025-09-29 |
| 40 |  MiniMax M2.7open ↗ | 45.1% | 55.6 | 2026-03-18 |
| 41 | Claude Opus 5medium | 44.9% | 64.6 | 2026-07-24 |
| 42 | Claude Fable 5max | 44.5% | 69.3 | 2026-06-09 |
| 43 | GPT-5-Codex | 44.3% | – | 2025-09-15 |
| 44 | Claude Fable 5.1low | 43.3% | 65.0 | 2026-09-01 |
| 45 |  Kimi K2.5open ↗ | 43.2% | 54.5 | 2026-01-27 |
| 46 | MiniMax M2.5open ↗ | 42.7% | 53.6 | 2026-02-12 |
| 47 | GLM 5.3maxopen ↗ | 41.8% | 63.5 | 2026-08-14 |
| 48 | Claude Opus 4.1 | 38.0% | 51.5 | 2025-08-05 |
| 49 | Grok 4.7xhigh | 37.6% | 61.6 | 2026-09-21 |
| 50 | GPT-5.6 Solmax | 37.3% | 68.0 | 2026-06-26 |
| 51 | MiniMax-M2.1open ↗ | 36.6% | 53.8 | 2025-12-23 |
| 52 | GLM 5.3 Flashno reasoningopen ↗ | 35.8% | – | 2026-08-26 |
| 53 | Kimi K2 Thinkingthinkingopen ↗ | 35.7% | 52.4 | 2025-11-06 |
| 54 | Claude Haiku 4.5 | 35.5% | 45.3 | 2025-10-15 |
| 55 | Claude Opus 5low | 34.9% | 61.3 | 2026-07-24 |
| 56 | GPT-5 mini | 34.8% | 56.5 | 2025-08-07 |
| 57 | GLM 4.7open ↗ | 33.4% | 51.9 | 2025-12-22 |
| 58 | Gemini 2.5 Pro | 32.6% | 53.9 | 2025-06-17 |
| 59 | GPT-5 minimedium | 31.9% | 53.8 | 2025-08-07 |
| 60 | MiniMax-M2open ↗ | 30.0% | 50.8 | 2025-10-27 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “Terminal-Bench leaderboard”, data as of 25 Oct 2026, https://www.benchleader.com/benchmarks/terminal_bench.
Terminal-Bench: questions
- What does Terminal-Bench measure?
- Terminal-Bench gives an agent a shell in a container and a task such as building a project from source, recovering a corrupted file system, or setting up a service, then checks the end state with tests. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads Terminal-Bench?
- GPT-5.5 leads Terminal-Bench with 84.7% as of 25 Oct 2026, ahead of GPT-5.4 at 81.8%.
- How many models have Terminal-Bench results?
- 84 model configurations have a Terminal-Bench result on BenchLeader, all taken from Terminal-Bench via Epoch AI Benchmarking Hub.
- Who runs Terminal-Bench and how often is it updated?
- Terminal-Bench is published by Terminal-Bench. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does Terminal-Bench count toward the BenchLeader Index?
- Yes. Terminal-Bench contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.