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.
- Made by
- Stanford University and Laude Institute
- Released
- May 2025
- Kind
- Terminal tasks
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Agent task
- Tasks
- All public
- Strongest model tested
- GPT-5 #70 on the capability index
- Last in a launch post
- December 2025, dropped for a newer version
- Last model added
- December 2025
- Scores refreshed
- November 12th, 2026
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/ .
GPT-5 leads with 49.6% among models released by October 2025
Best score per model from tbench.ai, as Epoch AI copied it. We don’t read that board ourselves yet. Where a model ran with several settings, we keep its best.
Claude Opus 4 leads with 50%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
The newest model with a score here came out in October 2025, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Fable 5.2
-
Gemini 4 Carbon - GPT-6 Astra
- Claude Opus 5.5
- 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%
Source: Tbench.ai, via Epoch AI (CC BY 4.0)
- 50.0 Claude Opus 4
- 50.0 Claude Sonnet 4.5
- 49.6 GPT-5
- 47.1 Kimi K2 Thinking
- 46.3 MiniMax-M2
- 45.3
UI-TARS-2 - 44.5 Kimi K2
- 44.3 GPT-5-Codex
- 43.3 Claude Opus 4.1
- 41.3 Claude Sonnet 4
- 41.3
Nova 2.0 Pro Preview - 41.0 Claude Haiku 4.5
- 40.5 GLM-4.6
- 39.5
LongCat-Flash-Chat - 37.7
DeepSeek-V3.2-Exp - 37.5 Qwen3-Coder-480B-A35B
- 37.5 GLM-4.5
- 36.7 DeepSeek-V3.1-Terminus
- 34.8 GPT-5 mini
- 34.6
Tencent HY 2.0 Think - 32.6 Gemini 2.5 Pro
- 32.5 Nova 2 Lite
- 31.3 DeepSeek-V3.1
- 30.0 Kimi K2
- 30.0 GLM-4.5-Air
- 27.2 Grok 4
- 26.3
Code World Model (CWM) - 25.8 grok-code-fast-1
- 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 Claude Opus 4 50%
- 2 Claude Sonnet 4.5 50%
- 3 GPT-5 unknown thinking 49.6%
- 4 Kimi K2 Thinking 47.1%
- 5 MiniMax-M2 46.3%
- 6 UI-TARS-2 45.3%
- 7 Kimi K2 Sep 2025 44.5%
- 8 GPT-5-Codex 44.3%
- 9 Claude Opus 4.1 43.3%
- 10 Claude Sonnet 4 41.3%
- 11 Nova 2.0 Pro Preview 41.3%
- 12 Claude Haiku 4.5 41%
- 13 GLM-4.6 40.5%
- 14 LongCat-Flash-Chat 39.5%
- 15 DeepSeek-V3.2-Exp 37.7%
- 16 Qwen3-Coder-480B-A35B 37.5%
- 17 GLM-4.5 37.5%
- 18 DeepSeek-V3.1-Terminus 36.7%
- 19 GPT-5 mini unknown thinking 34.8%
- 20 Tencent HY 2.0 Think 34.6%
- 21 Gemini 2.5 Pro Jun 2025 32.6%
- 22 Nova 2 Lite 32.5%
- 23 DeepSeek-V3.1 31.3%
- 24 Kimi K2 Jul 2025 30%
- 25 GLM-4.5-Air 30%
- 26 Grok 4 27.2%
- 27 Code World Model (CWM) 26.3%
- 28 grok-code-fast-1 25.8%
- 29 GPT-5 nano unknown thinking 21.8%
- 30 gpt-oss-120b unknown thinking 18.7%
- 31 Gemini 2.5 Flash Jun 2025 17.1%
- 32 Gemini 2.5 Flash Sep 2025 16.9%
- 33 gpt-oss-20b 3.37%
Sources: Labs’ model cards and tbench.ai, via Epoch AI (CC BY 4.0)
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-5 | 49.6% | unknown thinking, Codex CLI | Aug 2025 | tbench.ai, via Epoch AI |
| 2 | Claude Sonnet 4.5 | 46.5% | unknown thinking, CAMEL-AI | Sep 2025 | tbench.ai, via Epoch AI |
| 3 | GPT-5-Codex | 44.3% | Codex CLI | Sep 2025 | tbench.ai, via Epoch AI |
| 4 | Claude Opus 4.1 | 38% | Terminus 2 | Aug 2025 | Terminal-Bench v2 Leaderboard, via Epoch AI |
| 5 | Claude Haiku 4.5 | 35.5% | unknown thinking, Goose | Oct 2025 | tbench.ai, via Epoch AI |
| 6 | GPT-5 mini | 34.8% | unknown thinking, spoox-o-m | Aug 2025 | tbench.ai, via Epoch AI |
| 7 | Gemini 2.5 Pro Jun 2025 | 32.6% | Terminus 2 | Jun 2025 | tbench.ai, via Epoch AI |
| 8 | MiniMax-M2 | 30% | Terminus 2 | Oct 2025 | tbench.ai, via Epoch AI |
| 9 | Kimi K2 Jul 2025 | 27.8% | Terminus 2 | Jul 2025 | tbench.ai, via Epoch AI |
| 10 | Grok 4 | 27.2% | OpenHands | Jul 2025 | tbench.ai, via Epoch AI |
| 11 | Qwen3-Coder-480B-A35B | 27.2% | Dakou Agent | Jul 2025 | tbench.ai, via Epoch AI |
| 12 | grok-code-fast-1 | 25.8% | Mini-SWE-Agent | Aug 2025 | tbench.ai, via Epoch AI |
| 13 | GLM-4.6 | 24.5% | Terminus 2 | Sep 2025 | tbench.ai, via Epoch AI |
| 14 | GPT-5 nano | 21.8% | unknown thinking, spoox-o-m | Aug 2025 | tbench.ai, via Epoch AI |
| 15 | gpt-oss-120b | 18.7% | unknown thinking, Terminus 2 | Aug 2025 | tbench.ai, via Epoch AI |
| 16 | Gemini 2.5 Flash Jun 2025 | 17.1% | Mini-SWE-Agent | Jun 2025 | tbench.ai, via Epoch AI |
| 17 | Gemini 2.5 Flash Sep 2025 | 16.9% | Terminus 2 | Sep 2025 | Terminal-Bench v2 Leaderboard, via Epoch AI |
| 18 | gpt-oss-20b | 3.37% | Mini-SWE-Agent | Aug 2025 | tbench.ai, via Epoch AI |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Opus 4 | 50% | May 2025 | Lab model card | |
| 2 | Claude Sonnet 4.5 | 50% | Sep 2025 | Lab model card | |
| 3 | GPT-5 | 49.6% | unknown thinking, Codex CLI | Aug 2025 | tbench.ai, via Epoch AI |
| 4 | Kimi K2 Thinking | 47.1% | Nov 2025 | Lab model card | |
| 5 | MiniMax-M2 | 46.3% | Oct 2025 | Lab model card | |
| 6 | UI-TARS-2 | 45.3% | Sep 2025 | Lab model card | |
| 7 | Kimi K2 Sep 2025 | 44.5% | Sep 2025 | Lab model card | |
| 8 | GPT-5-Codex | 44.3% | Codex CLI | Sep 2025 | tbench.ai, via Epoch AI |
| 9 | Claude Opus 4.1 | 43.3% | Aug 2025 | Lab model card | |
| 10 | Claude Sonnet 4 | 41.3% | May 2025 | Lab model card | |
| 11 | Nova 2.0 Pro Preview | 41.3% | Dec 2025 | Lab model card | |
| 12 | Claude Haiku 4.5 | 41% | Oct 2025 | Lab model card | |
| 13 | GLM-4.6 | 40.5% | Sep 2025 | Lab model card | |
| 14 | LongCat-Flash-Chat | 39.5% | Aug 2025 | Lab model card | |
| 15 | DeepSeek-V3.2-Exp | 37.7% | Sep 2025 | Lab model card | |
| 16 | Qwen3-Coder-480B-A35B | 37.5% | Jul 2025 | Lab model card | |
| 17 | GLM-4.5 | 37.5% | Jul 2025 | Lab model card | |
| 18 | DeepSeek-V3.1-Terminus | 36.7% | Sep 2025 | Lab model card | |
| 19 | GPT-5 mini | 34.8% | unknown thinking, spoox-o-m | Aug 2025 | tbench.ai, via Epoch AI |
| 20 | Tencent HY 2.0 Think | 34.6% | Dec 2025 | Lab model card | |
| 21 | Gemini 2.5 Pro Jun 2025 | 32.6% | Terminus 2 | Jun 2025 | tbench.ai, via Epoch AI |
| 22 | Nova 2 Lite | 32.5% | Dec 2025 | Lab model card | |
| 23 | DeepSeek-V3.1 | 31.3% | Aug 2025 | Lab model card | |
| 24 | Kimi K2 Jul 2025 | 30% | Jul 2025 | Lab model card | |
| 25 | GLM-4.5-Air | 30% | Aug 2025 | Lab model card | |
| 26 | Grok 4 | 27.2% | OpenHands | Jul 2025 | tbench.ai, via Epoch AI |
| 27 | Code World Model (CWM) | 26.3% | Sep 2025 | Lab model card | |
| 28 | grok-code-fast-1 | 25.8% | Mini-SWE-Agent | Aug 2025 | tbench.ai, via Epoch AI |
| 29 | GPT-5 nano | 21.8% | unknown thinking, spoox-o-m | Aug 2025 | tbench.ai, via Epoch AI |
| 30 | gpt-oss-120b | 18.7% | unknown thinking, Terminus 2 | Aug 2025 | tbench.ai, via Epoch AI |
| 31 | Gemini 2.5 Flash Jun 2025 | 17.1% | Mini-SWE-Agent | Jun 2025 | tbench.ai, via Epoch AI |
| 32 | Gemini 2.5 Flash Sep 2025 | 16.9% | Terminus 2 | Sep 2025 | Terminal-Bench v2 Leaderboard, via Epoch AI |
| 33 | gpt-oss-20b | 3.37% | Mini-SWE-Agent | Aug 2025 | tbench.ai, via Epoch AI |
Labs report scores for 22 models in their cards. Switch the scores above to rank them together.
3 models have held number one on tbench.ai since its release in May 2025 Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for tbench.ai, which Epoch AI copied and what labs claim in their cards.
Sources: Tbench.ai, via Epoch AI (CC BY 4.0) and labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Other boards, via Epoch AI (CC BY 4.0) and labs’ model cards
Labs headline it less often now, last in December 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 |
|---|---|---|---|
| Dec 5, 2025 | Tencent HY 2.0 Think | 34.6 | Terminal Bench |
| Dec 2, 2025 | Nova 2 Lite | 32.5 | Terminal-Bench 1.0 accuracy |
| Dec 2, 2025 | Nova 2.0 Pro Preview | 41.3 | Terminal-Bench 1.0 accuracy |
| Nov 6, 2025 | Kimi K2 Thinking | 47.1 | Terminal-Bench, w/ simulated tools (JSON) |
| Oct 27, 2025 | MiniMax-M2 | 46.3 | Terminal-Bench |
| Oct 15, 2025 | Claude Haiku 4.5 | 41 | Agentic terminal coding Terminal-Bench |
| Sep 30, 2025 | GLM-4.6 | 40.5 | Terminal-Bench |
| Sep 29, 2025 | DeepSeek-V3.2-Exp | 37.7 | Terminal-bench |
| Sep 29, 2025 | Claude Sonnet 4.5 | 50 | Agentic terminal coding Terminal-Bench (Terminus 2) |
| Sep 22, 2025 | DeepSeek-V3.1-Terminus | 36.7 | Terminal-bench |
| Sep 5, 2025 | Kimi K2 | 44.5 | Terminal-Bench, ACC |
| Sep 2, 2025 | UI-TARS-2 | 45.3 | TB (Terminal Bench) with GUI-SDK, 75 of 80 tasks |
| Aug 29, 2025 | LongCat-Flash-Chat | 39.5 | TerminalBench(acc) |
| Aug 21, 2025 | DeepSeek-V3.1 | 31.3 | Terminal-bench (Terminus 1 framework) |
| Aug 5, 2025 | Claude Opus 4.1 | 43.3 | Agentic terminal coding Terminal-Bench (Terminus 1) |
| Jul 28, 2025 | GLM-4.5 | 37.5 | SWE-Terminal-Bench (Terminal Bench) |
| Jul 22, 2025 | Qwen3-Coder-480B-A35B | 37.5 | Terminal-Bench |
| Jul 11, 2025 | Kimi K2 | 30 | TerminalBench, Inhouse Framework (Acc) |
| May 22, 2025 | Claude Sonnet 4 | 41.3 | Agentic terminal coding Terminal-bench (parallel test-time compute) |
| May 22, 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 |
|---|---|---|---|---|
| Dec 2, 2025 | Nova 2 Lite technical report | 32.5 | Terminal-Bench 1.0 accuracy, 80 tasks (Table 4) | Capabilities |
| Dec 2, 2025 | Nova 2.0 Pro Preview technical report | 41.3 | Terminal-Bench 1.0 accuracy, 80 tasks (Table 4) | Capabilities |
| Sep 30, 2025 | Code World Model (CWM) technical report | 26.3 | Terminal-Bench, Terminus 1 agent, 50 turns, 26.25 in text (Table 7) | Capabilities |
| Sep 29, 2025 | DeepSeek-V3.2-Exp technical report | 37.7 | Terminal-bench (Terminus 1 framework), DeepSeek-V3.2-Exp (Table 1) | Capabilities |
| Sep 1, 2025 | LongCat-Flash-Chat technical report | 39.5 | TerminalBench (acc), Terminus framework (Table 3) | Capabilities |
| Aug 8, 2025 | GLM-4.5 technical report | 37.5 | Terminal-Bench, Terminus framework (Table 5) | Capabilities |
| Aug 8, 2025 | GLM-4.5-Air technical report | 30 | Terminal-Bench, Terminus framework (Table 5) | Capabilities |
| Jul 21, 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.
Sources: Official leaderboards and boards copied by Epoch AI (CC BY 4.0)
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%
- Terminal-Bench 2.1 Extended Not in model cards Agents complete long command-line tasks in containers, scored by verification tests. Mercor adds 99 new expert-written tasks to the 89 public ones and reports both scores. Muse Wonder 54%
- o11y-bench Not in model cards Agents query metrics, logs and traces in a live Grafana stack, investigate incidents and fix dashboards. Grading checks claims against facts read from the same environment. Claude Opus 5.5 93.8%
- 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%