Terminal-Bench
Agents complete about 80 hard tasks in sandboxed Linux terminals, such as compiling code or configuring servers. Test scripts check the final container state.
Terminal-Bench 2.0 replaced this first release in November 2025 with 89 harder tasks, and later versions followed.
In the authors’ words
AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 2.0: a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 65% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work at https://www.tbench.ai/ .
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
GPT-5 leads with 49.6% among models released by Oct 2025
Best score per model in Epoch AI’s data. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in Oct 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6 Astra
- Claude Opus 5.5
- GPT-6.1 Terra
- GPT-6.1 Sol
- and 16 more
- 49.6 GPT-5
- 46.5 Claude Sonnet 4.5
- 44.3 GPT-5-Codex
- 38.0 Claude Opus 4.1
- 35.5 Claude Haiku 4.5
- 34.8 GPT-5 mini
- 32.6
Gemini 2.5 Pro - 30.0
MiniMax-M2 - 27.8 Kimi K2
- 27.2 Grok 4
- 27.2
Qwen3-Coder-480B-A35B - 25.8 grok-code-fast-1
- 24.5 GLM-4.6
- 21.8 GPT-5 nano
- 18.7 gpt-oss-120b
- 17.1 Gemini 2.5 Flash
- 16.9 Gemini 2.5 Flash
- 3.37 gpt-oss-20b
- 1 GPT-5 unknown thinking 49.6%
- 2 Claude Sonnet 4.5 unknown thinking 46.5%
- 3 GPT-5-Codex 44.3%
- 4 Claude Opus 4.1 38%
- 5 Claude Haiku 4.5 unknown thinking 35.5%
- 6 GPT-5 mini unknown thinking 34.8%
- 7 Gemini 2.5 Pro Jun 2025 32.6%
- 8 MiniMax-M2 30%
- 9 Kimi K2 Jul 2025 27.8%
- 10 Grok 4 27.2%
- 11 Qwen3-Coder-480B-A35B 27.2%
- 12 grok-code-fast-1 25.8%
- 13 GLM-4.6 24.5%
- 14 GPT-5 nano unknown thinking 21.8%
- 15 gpt-oss-120b unknown thinking 18.7%
- 16 Gemini 2.5 Flash Jun 2025 17.1%
- 17 Gemini 2.5 Flash Sep 2025 16.9%
- 18 gpt-oss-20b 3.37%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5 | 49.6% | unknown thinking, Codex CLI | Aug 2025 | Epoch AI |
| 2 | Claude Sonnet 4.5 | 46.5% | unknown thinking, CAMEL-AI | Sep 2025 | Epoch AI |
| 3 | GPT-5-Codex | 44.3% | Codex CLI | Sep 2025 | Epoch AI |
| 4 | Claude Opus 4.1 | 38% | Terminus 2 | Aug 2025 | Epoch AI |
| 5 | Claude Haiku 4.5 | 35.5% | unknown thinking, Goose | Oct 2025 | Epoch AI |
| 6 | GPT-5 mini | 34.8% | unknown thinking, spoox-o-m | Aug 2025 | Epoch AI |
| 7 | Gemini 2.5 Pro Jun 2025 | 32.6% | Terminus 2 | Jun 2025 | Epoch AI |
| 8 | MiniMax-M2 | 30% | Terminus 2 | Oct 2025 | Epoch AI |
| 9 | Kimi K2 Jul 2025 | 27.8% | Terminus 2 | Jul 2025 | Epoch AI |
| 10 | Grok 4 | 27.2% | OpenHands | Jul 2025 | Epoch AI |
| 11 | Qwen3-Coder-480B-A35B | 27.2% | Dakou Agent | Jul 2025 | Epoch AI |
| 12 | grok-code-fast-1 | 25.8% | Mini-SWE-Agent | Aug 2025 | Epoch AI |
| 13 | GLM-4.6 | 24.5% | Terminus 2 | Sep 2025 | Epoch AI |
| 14 | GPT-5 nano | 21.8% | unknown thinking, spoox-o-m | Aug 2025 | Epoch AI |
| 15 | gpt-oss-120b | 18.7% | unknown thinking, Terminus 2 | Aug 2025 | Epoch AI |
| 16 | Gemini 2.5 Flash Jun 2025 | 17.1% | Mini-SWE-Agent | Jun 2025 | Epoch AI |
| 17 | Gemini 2.5 Flash Sep 2025 | 16.9% | Terminus 2 | Sep 2025 | Epoch AI |
| 18 | gpt-oss-20b | 3.37% | Mini-SWE-Agent | Aug 2025 | Epoch AI |
Labs report scores for 22 models in their cards. We plot every source in the charts below.
3 models have held number one since its release in May 2025
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for Epoch AI’s runs and what labs claim in their cards.
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs headline it less often now, last in Dec 2025
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 5 Dec 2025 |  Tencent HY 2.0 Think | 34.6 | Terminal Bench |
| 2 Dec 2025 |  Nova 2 Lite | 32.5 | Terminal-Bench 1.0 accuracy |
| 2 Dec 2025 | Nova 2.0 Pro Preview | 41.3 | Terminal-Bench 1.0 accuracy |
| 6 Nov 2025 | Kimi K2 Thinking | 47.1 | Terminal-Bench, w/ simulated tools (JSON) |
| 27 Oct 2025 | MiniMax-M2 | 46.3 | Terminal-Bench |
| 15 Oct 2025 | Claude Haiku 4.5 | 41 | Agentic terminal coding Terminal-Bench |
| 30 Sep 2025 | GLM-4.6 | 40.5 | Terminal-Bench |
| 29 Sep 2025 |  DeepSeek-V3.2-Exp | 37.7 | Terminal-bench |
| 29 Sep 2025 | Claude Sonnet 4.5 | 50 | Agentic terminal coding Terminal-Bench (Terminus 2) |
| 22 Sep 2025 | DeepSeek-V3.1-Terminus | 36.7 | Terminal-bench |
| 5 Sep 2025 | Kimi K2 | 44.5 | Terminal-Bench, ACC |
| 2 Sep 2025 |  UI-TARS-2 | 45.3 | TB (Terminal Bench) with GUI-SDK, 75 of 80 tasks |
| 29 Aug 2025 |  LongCat-Flash-Chat | 39.5 | TerminalBench(acc) |
| 21 Aug 2025 | DeepSeek-V3.1 | 31.3 | Terminal-bench (Terminus 1 framework) |
| 5 Aug 2025 | Claude Opus 4.1 | 43.3 | Agentic terminal coding Terminal-Bench (Terminus 1) |
| 28 Jul 2025 | GLM-4.5 | 37.5 | SWE-Terminal-Bench (Terminal Bench) |
| 22 Jul 2025 | Qwen3-Coder-480B-A35B | 37.5 | Terminal-Bench |
| 11 Jul 2025 | Kimi K2 | 30 | TerminalBench, Inhouse Framework (Acc) |
| 22 May 2025 | Claude Sonnet 4 | 41.3 | Agentic terminal coding Terminal-bench (parallel test-time compute) |
| 22 May 2025 | Claude Opus 4 | 50 | Agentic terminal coding Terminal-bench (parallel test-time compute) |
8 system cards report it, from 6 labs
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 2 Dec 2025 | Nova 2 Lite technical report | 32.5 | Terminal-Bench 1.0 accuracy, 80 tasks (Table 4) | Capabilities |
| 2 Dec 2025 | Nova 2.0 Pro Preview technical report | 41.3 | Terminal-Bench 1.0 accuracy, 80 tasks (Table 4) | Capabilities |
| 30 Sep 2025 |  Code World Model (CWM) technical report | 26.3 | Terminal-Bench, Terminus 1 agent, 50 turns, 26.25 in text (Table 7) | Capabilities |
| 29 Sep 2025 | DeepSeek-V3.2-Exp technical report | 37.7 | Terminal-bench (Terminus 1 framework), DeepSeek-V3.2-Exp (Table 1) | Capabilities |
| 1 Sep 2025 | LongCat-Flash-Chat technical report | 39.5 | TerminalBench (acc), Terminus framework (Table 3) | Capabilities |
| 8 Aug 2025 | GLM-4.5 technical report | 37.5 | Terminal-Bench, Terminus framework (Table 5) | Capabilities |
| 8 Aug 2025 | GLM-4.5-Air technical report | 30 | Terminal-Bench, Terminus framework (Table 5) | Capabilities |
| 21 Jul 2025 | Kimi K2 technical report | 30 | Terminal Bench In-House framework (Acc.). Terminus framework 25.0 (Table 3) | Capabilities |
Versions
We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.
More in Terminal agents
- SkillsBench In use 86-87 containerized expert tasks in 11 domains, run with and without Agent Skills. Verifier tests give pass rate and skill lift. GPT-5.5 67.3%
- Terminal-Bench Hard In use Agents complete 44 hard tasks from Terminal-Bench Core in the Terminus 2 harness. Artificial Analysis scores pass@1 averaged over 3 repeats. Gemma 4 31B IT 36%
- Harbor-Index In use Agents solve about 80 hard tasks drawn from 29 benchmarks run through Harbor adapters, such as HLE, GSO and SWE-bench Pro. Each task keeps its source verifier. GPT-5.5 28.1%
- CompileBench Not in model cards Agents build real open-source projects through dependency problems, legacy toolchains and complex build systems, for example compiling 22-year-old code. Claude Opus 4.1 100%
- Long-Horizon-Terminal-Bench 2.0 Not in model cards Agents work for up to 3 hours per task on 77 container tasks, including 30 tasks that rebuild a scientific tool from its outputs. Hidden verifiers give partial-credit rewards. GPT-6 Astra 69.2%
- ITBench Not in model cards Agents handle IT automation tasks in Kubernetes environments: site reliability incidents, compliance and security checks, and cloud cost work. Gemini 3 Pro 58.3%