AI Rankings
Terminal-Bench Leaderboard: AI Models Ranked by Score
Data updated
What Terminal-Bench measures
Tasks done in a computer terminal, such as setting up software or fixing a broken build, checked by automatic tests. Scores here use the benchmark's standard reference agent so models are compared on the same footing.
Scores are percentages and higher is better. Results come from Epoch AI's open benchmark data; we do not run the test ourselves.
Top 20 at a glance
Full ranking
| Rank and model | Terminal-Bench score | Epoch index | Price per 1M tokens (in / out) | Released |
|---|---|---|---|---|
1GPT-5.3 CodexOpenAI | 64.7% | 157.1 | $1.75 / $14.00 | 2026-02-05 |
2Claude Opus 4.6Anthropic | 62.9% | 155.7 | $5.00 / $25.00 | 2026-02-05 |
3Claude Opus 4.5Anthropic | 57.8% | 150.1 | $5.00 / $25.00 | 2025-11-24 |
4Gemini 3 ProGoogle | 56.9% | 153.1 | $2.00 / $12.00 | 2025-11-18 |
5GPT-5.2OpenAI | 54.0% | 153.7 | $1.75 / $14.00 | 2025-12-11 |
6GLM-5Z.ai · Open weights | 52.4% | 145.8 | $1.00 / $3.20 | 2026-02-11 |
7Gemini 3 FlashGoogle | 51.7% | 151.9 | $0.50 / $3.00 | 2025-12-17 |
8GPT-5.1OpenAI | 47.6% | 149.8 | $1.25 / $10.00 | 2025-11-13 |
9Kimi K2.5Moonshot AI · Open weights | 43.2% | 148.1 | $0.60 / $3.00 | 2026-01-27 |
10Claude Sonnet 4.5Anthropic | 42.8% | 146.7 | $3.00 / $15.00 | 2025-09-29 |
11MiniMax-M2.5MiniMax · Open weights | 42.3% | 146.6 | $0.30 / $1.20 | 2026-02-12 |
12DeepSeek-V3.2DeepSeek · Open weights | 39.5% | 146.1 | $0.28 / $0.40 | 2025-12-01 |
13Claude Opus 4.1Anthropic | 38.0% | 144.1 | $15.00 / $75.00 | 2025-08-05 |
14Kimi K2 ThinkingMoonshot AI · Open weights | 35.7% | 145.8 | $0.60 / $2.50 | 2025-11-06 |
15GPT-5OpenAI | 35.2% | 150.0 | $1.25 / $10.00 | 2025-08-07 |
16GLM-4.7Z.ai · Open weights | 33.4% | 143.6 | $0.60 / $2.20 | 2025-12-22 |
17Gemini 2.5 Pro (Jun 2025)Google | 32.6% | 145.1 | $1.25 / $10.00 | 2025-06-05 |
18Claude Haiku 4.5Anthropic | 28.3% | 142.2 | $1.00 / $5.00 | 2025-10-15 |
19Kimi K2 (Jul 2025)Moonshot AI · Open weights | 27.8% | 140.0 | $0.55 / $2.20 | 2025-07-12 |
20GPT-5 miniOpenAI | 24.0% | 145.3 | $0.25 / $2.00 | 2025-08-07 |
21Grok 4xAI | 23.1% | 146.2 | $3.00 / $15.00 | 2025-07-09 |
22gpt-oss-120bOpenAI · Open weights | 18.7% | 139.8 | $0.15 / $0.60 | 2025-08-05 |
23Gemini 2.5 Flash (Sep 2025)Google | 16.9% | 142.9 | $0.30 / $2.50 | 2025-09-25 |
24GPT-5 nanoOpenAI | 7.9% | 139.2 | $0.05 / $0.40 | 2025-08-07 |
25gpt-oss-20bOpenAI · Open weights | 3.1% | 137.7 | $0.06 / $0.24 | 2025-08-05 |
How to read this leaderboard
Scores are percentages and higher is better. For each model we show the best result Epoch lists across its effort settings, the same rule for every model. A gap of a point or two is usually within the normal noise of a test, so treat close scores as a tie. Newer models can be missing when nobody has run the test on them yet, so check the release dates. A benchmark measures one kind of skill: use it together with the overall leaderboard rather than on its own.
Questions people ask
What does Terminal-Bench measure?
Tasks done in a computer terminal, such as setting up software or fixing a broken build, checked by automatic tests. Scores here use the benchmark's standard reference agent so models are compared on the same footing.
Who ran these tests?
The results come from Epoch AI's open benchmark data. It combines runs by Epoch with results published by other groups, and Epoch's own pages say which is which. We do not run tests ourselves.
Does a high score mean the model is best for me?
Not by itself. Terminal-Bench checks one kind of ability. A model can lead here and still be slower, costlier or a poor fit for your work. Compare it on the overall leaderboard and try it on your own tasks.
Why is a model I know missing?
A model appears only when an open source has published results for it. New models can take days or weeks to be measured, and models with no published score are left out rather than guessed.
Where does the data come from?
Scores come from Epoch AI's open benchmark data (licensed CC BY 4.0). We match models across sources, rank them and show them in charts, without changing the scores. Epoch AI does not endorse this site. Prices and context windows come from the open-source LiteLLM price list (MIT licence), which records each model maker's own price page. Prices change often, so confirm on the maker's site before you buy.
How often is this leaderboard updated?
We check our open data sources every day and show the date of the data at the top of the page. A model only moves when its source publishes new results, so a number can stay the same for weeks.
More leaderboards
- Overall intelligenceWhich AI model is the smartest overall?Leads: Claude Opus 5.5 (168.8)270 models ranked
- Human preferenceWhich AI do people prefer in blind tests?Leads: Gemini 4 Argon (1517)166 models ranked
- CodingWhich AI model is best for coding?Leads: Gemini 4 Argon (1552)166 models ranked
- Web developmentWhich AI model builds the best websites and apps?Leads: Claude Opus 5.5 (1807)85 models ranked
- Image understandingWhich AI model understands images best?Leads: Claude Fable 5 (1309)85 models ranked
- Open-weights modelsWhich open-source AI model is best?Leads: Kimi K3 (157.9)123 models ranked
- Cheapest strong modelsWhat is the cheapest AI model that is still good?Cheapest: gpt-oss-20b ($0.10)119 models ranked
- Best valueWhich AI models give the most intelligence for the money?Line starts at: gpt-oss-20b ($0.10)7 models ranked
- Context windowWhich AI model can read the most text at once?Leads: GPT-5.4 (1.1M tokens)193 models ranked
- GPQA DiamondWhich AI model scores highest on GPQA Diamond?Leads: Gemini 4 Argon (96.4%)183 models ranked
- Humanity's Last ExamWhich AI model scores highest on Humanity's Last Exam?Leads: GPT-6 Astra (54.8%)41 models ranked
- SWE-bench VerifiedWhich AI model scores highest on SWE-bench Verified?Leads: Claude Opus 4.7 (83.5%)32 models ranked
- Aider PolyglotWhich AI model scores highest on Aider Polyglot?Leads: GPT-5 (88.0%)46 models ranked
- FrontierMath (Tiers 1-3)Which AI model scores highest on FrontierMath (Tiers 1-3)?Leads: GPT-6 Astra (93.7%)87 models ranked
- ARC-AGI-2Which AI model scores highest on ARC-AGI-2?Leads: GPT-6 Astra (95.0%)86 models ranked
- SimpleQA VerifiedWhich AI model scores highest on SimpleQA Verified?Leads: Gemini 4 Argon (76.0%)82 models ranked
- SimpleBenchWhich AI model scores highest on SimpleBench?Leads: Claude Fable 5 (81.9%)80 models ranked
- DeepSWEWhich AI model scores highest on DeepSWE?Leads: GPT-6 Astra (74.1%)26 models ranked
Where this data comes from
- Epoch AI: Benchmark data: Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks (CC BY 4.0). We match models across sources, rank them and show them in charts. Scores are not changed. Epoch AI does not endorse this site.
- LiteLLM model price list: Prices and context windows: LiteLLM model price list (github.com/BerriAI/litellm, MIT licence). We use the model maker's own price entry when there is one. Prices change often, so always confirm on the maker's site before you buy.
Newest data on this page: 26 Oct 2026. Prices are list prices and change often, so confirm on the maker's site before you buy. Sponsors never change a score or an order. How we calculate this · Report a mistake