Cleus
Start free
  1. Cleus
  2. AI benchmarks

AI model benchmarkscompared and explained

How today’s top AI models really compare on reasoning, maths, coding, facts and long tasks, with every test explained in plain English so you know what a score means before you trust it.

18 models · 11 benchmarks · our own hands-on tests · updated November 9, 2026

Find the right AI for youTest two models on your promptMost uncensored models
Top model overall
GPT-6 Astra
Builds the best websites
Claude Opus 5.5
Best value at the top
Claude Haiku 5.5
Best open-weights model
DeepSeek V4.1 Flash
01

Leaderboard

Overall rank uses Epoch AI’s capability index, which combines dozens of tests into one score. Switch tabs to rank each skill on its own.

#ContextReleased
1GPT-6 AstraOpenAI169.1$10 / $501.05MSep 2026
2GPT-6.1 SolOpenAI168.9$2 / $101.05MSep 2026
3Claude Opus 5.5Anthropic168.4$4 / $201MSep 2026
4GPT-6.1 TerraOpenAI167.8$1 / $51.05MOct 2026
5Claude Fable 5.1Anthropic166.9$10 / $501MSep 2026
6Gemini 4 ArgonGoogle166.4$2 / $101.05MSep 2026
7Claude Sonnet 5.5Anthropic166.2$2 / $101MSep 2026
8Muse WonderMeta162.0$4 / $201.05MOct 2026
9Claude Haiku 5.5Anthropic160.2$0.10 / $0.501MOct 2026
10DeepSeek V4.1 FlashDeepSeekopen159.3$0.006 / $0.501.05MOct 2026
11Gemini 3.8 FlashGoogle158.6$0.82 / $4.131.05MSep 2026
12Kimi K3Moonshot AIopen158.5$0.19 / $161.05MJul 2026
13Grok 4.6xAI157.9$2 / $6500KAug 2026
14GLM-5.3Z.aiopen156.9$0.070 / $1.401.05MAug 2026
15Grok 4.7xAI156.1$2 / $6500KSep 2026
16Muse Spark OpenMetaopen155.4$1.25 / $4.251.05MOct 2026
17Mistral Large 4Mistral AIopen154.7$0.70 / $2.101.05MOct 2026
18Kimi K3.1 ProMoonshot AIopen—$2 / $101.05MOct 2026

Click a column to sort. “—” means the model hasn’t been run on that test yet. Highlighted cells are the best score in the column.

02

Try your own prompt

Benchmarks are averages. See how two models handle the thing you actually need, side by side, live.

Free: 3 side-by-side runs a day, more when you’re signed in. Answers are capped at about 250 words and run live through each model’s API. Prompts that ask for sexual content involving minors or mass-casualty weapons are never sent.

03

The race, and what it costs

Each lab’s best model over time, and how much capability each dollar buys.

Best model per lab

Epoch Capabilities Index since January 2025. Each step is a release that beat the lab’s previous best.

12513013514014515015516016517017520252026o1 (OpenAI): 141.16o3 (OpenAI): 146.99o3-pro (OpenAI): 147.41GPT-5 (OpenAI): 150GPT-5 Pro (OpenAI): 151.15GPT-5.2 Pro (OpenAI): 157.22GPT-5.4 Pro (OpenAI): 160.21GPT-5.5 Pro (OpenAI): 162.67GPT-5.6 Sol (OpenAI): 162.99GPT-6 Astra (OpenAI): 169.07Claude 3.5 Sonnet (October 2024) (Anthropic): 133.42Claude 3.7 Sonnet (Anthropic): 140.99Claude Opus 4 (Anthropic): 142.83Claude Opus 4.1 (Anthropic): 144.74Claude Sonnet 4.5 (Anthropic): 147.56Claude Opus 4.5 (Anthropic): 151.82Claude Opus 4.6 (Anthropic): 156.3Claude Opus 4.7 (Anthropic): 157.55Claude Opus 4.8 (Anthropic): 159.89Claude Fable 5 (Anthropic): 164.19Claude Opus 5 (Anthropic): 164.45Claude Fable 5.1 (Anthropic): 166.9Claude Opus 5.5 (Anthropic): 168.36Gemini 1.5 Pro (Sept 2024) (Google): 130.45Gemini 2.0 Pro (Google): 135.26Gemini 2.5 Pro (Mar 2025) (Google): 143.9Gemini 2.5 Pro (Jun 2025) (Google): 144.79Gemini 3 Pro (Google): 152.88Gemini 3.1 Pro (Google): 156.24Gemini 3.7 Flash (Google): 158.4Gemini 3.8 Flash (Google): 158.63Gemini 4 Argon (Google): 166.45Grok-2 (Dec 2024) (xAI): 129.8Grok 3 (xAI): 137.5Grok-3 mini (xAI): 139.89Grok 4 (xAI): 146.86Grok 4.20 (xAI): 152.27Grok 4.5 (xAI): 155.08Grok 4.6 (xAI): 157.92Llama 3.1-405B (Meta): 127.88Llama 4 Maverick (Meta): 132.13Muse Spark (Meta): 153.1Muse Spark 1.1 (Meta): 154.87Muse Spark 1.2 (Meta): 155.05Muse Spark 1.3 (Meta): 157.98Muse Wonder (Meta): 162.02DeepSeek-V3 (DeepSeek): 131.77DeepSeek-R1 (DeepSeek): 138.62DeepSeek-R1 (May 2025) (DeepSeek): 140.75DeepSeek-V3.1 (DeepSeek): 140.85DeepSeek-V3.2-Exp (DeepSeek): 144.41DeepSeek-V3.2 (DeepSeek): 146.22DeepSeek-V4-Pro (DeepSeek): 150.13DeepSeek V4 Flash 0731 (DeepSeek): 154.34DeepSeek V4 Pro 0813 (DeepSeek): 155.87DeepSeek V4.1 Flash (DeepSeek): 156.49DeepSeek V4.1 Flash 1017 (DeepSeek): 159.31Kimi K2 (Jul 2025) (Moonshot): 141.25Kimi K2 Thinking (Moonshot): 145.51Kimi K2.5 (Moonshot): 148.72Kimi K2.6 (Moonshot): 151.56Kimi K3 (Moonshot): 158.51Qwen2.5-72B (Alibaba): 127.29Qwen2.5-Max (Alibaba): 131.21QwQ-32B (Alibaba): 136.52Qwen3-235B-A22B (Alibaba): 138.44Qwen3-235B-A22B-Thinking (Jul 2025) (Alibaba): 143.2Qwen3.5 397B-A17B (Alibaba): 147.87Qwen 3.6 Plus (Alibaba): 148.8Qwen 3.6 Max (Preview) (Alibaba): 149.77Qwen3.7-Max (Alibaba): 154.5Qwen 3.8 Max (Alibaba): 157.43OpenAIAnthropicGoogleMetaDeepSeekMoonshotxAIAlibaba
Each line is a lab’s best model so far on Epoch AI’s overall capability index. A step up is a new release that beat the lab’s previous best.

Capability against price

Higher and further left is better: more capable for less money.

153157161165169$0.10$1$10$100Price per million tokens (3 in : 1 out, log scale) →Overall score (ECI) →GPT-6 Astra: ECI 169.07, $20 per M tokensGPT-6 AstraGPT-6.1 Sol: ECI 168.94, $4 per M tokensGPT-6.1 SolClaude Opus 5.5: ECI 168.36, $8 per M tokensClaude Opus 5.5GPT-6.1 Terra: ECI 167.82, $2 per M tokensGPT-6.1 TerraClaude Fable 5.1: ECI 166.9, $20 per M tokensClaude Fable 5.1Gemini 4 Argon: ECI 166.45, $4 per M tokensGemini 4 ArgonClaude Sonnet 5.5: ECI 166.24, $4 per M tokensClaude Sonnet 5.5Muse Wonder: ECI 162.02, $8 per M tokensMuse WonderClaude Haiku 5.5: ECI 160.22, $0.20 per M tokensClaude Haiku 5.5DeepSeek V4.1 Flash: ECI 159.31, $0.13 per M tokensDeepSeek V4.1 FlashGemini 3.8 Flash: ECI 158.63, $1.65 per M tokensGemini 3.8 FlashKimi K3: ECI 158.51, $4.14 per M tokensKimi K3Grok 4.6: ECI 157.92, $3 per M tokensGrok 4.6GLM-5.3: ECI 156.91, $0.40 per M tokensGLM-5.3Grok 4.7: ECI 156.07, $3 per M tokensGrok 4.7Muse Spark Open: ECI 155.42, $2 per M tokensMuse Spark OpenMistral Large 4: ECI 154.66, $1.05 per M tokensMistral Large 4
Up and to the left is better: more capable for less money. Hover or tap a dot to see the model.
04

Our hands-on tests

Benchmarks don’t show what a model refuses, how fast it feels or whether its code runs, so we gave all 18 the same prompts on November 9, 2026.

Answers without a lecture

Six legal but sensitive requests, answered in full with no unasked-for warnings.

  1. 1GPT-6.1 Terra6 / 6
  2. 2GPT-6 Astra5 / 6
  3. 3GPT-6.1 Sol5 / 6
  4. 4Claude Opus 5.55 / 6
  5. 5Claude Fable 5.15 / 6
  6. 6Claude Sonnet 5.55 / 6
  7. 7Muse Wonder5 / 6
  8. 8Gemini 3.8 Flash5 / 6
  9. 9Kimi K35 / 6
  10. 10Grok 4.65 / 6
  11. 11GLM-5.35 / 6
  12. 12Kimi K3.1 Pro5 / 6
  13. 13Claude Haiku 5.54 / 6
  14. 14DeepSeek V4.1 Flash4 / 6
  15. 15Grok 4.74 / 6
  16. 16Mistral Large 44 / 6
  17. 17Gemini 4 Argon4 / 6
  18. 18Muse Spark Open4 / 6

Time to first word

Typical wait before text appears. Lower is better; thinking time counts.

  1. 1Claude Haiku 5.51.1s
  2. 2Claude Sonnet 5.51.1s
  3. 3Claude Fable 5.12.4s
  4. 4GPT-6.1 Terra2.7s
  5. 5Claude Opus 5.52.9s
  6. 6GPT-6 Astra3.4s
  7. 7GPT-6.1 Sol3.6s
  8. 8Mistral Large 44.7s
  9. 9Muse Spark Open5.1s
  10. 10DeepSeek V4.1 Flash5.4s
  11. 11Gemini 3.8 Flash7.9s
  12. 12Kimi K38.7s
  13. 13GLM-5.310.2s
  14. 14Muse Wonder10.3s
  15. 15Kimi K3.1 Pro10.6s
  16. 16Gemini 4 Argon11.4s
  17. 17Grok 4.711.6s
  18. 18Grok 4.617.4s

Output speed

Tokens per second once it starts, one test day through OpenRouter.

  1. 1Muse Spark Open229/s
  2. 2Claude Haiku 5.5185/s
  3. 3Muse Wonder165/s
  4. 4Mistral Large 4160/s
  5. 5DeepSeek V4.1 Flash157/s
  6. 6Kimi K3152/s
  7. 7Gemini 3.8 Flash146/s
  8. 8GPT-6.1 Terra141/s
  9. 9Claude Sonnet 5.5132/s
  10. 10GLM-5.3115/s
  11. 11Kimi K3.1 Pro114/s
  12. 12Claude Opus 5.5104/s
  13. 13GPT-6 Astra93/s
  14. 14Claude Fable 5.169/s
  15. 15Gemini 4 Argon67/s
  16. 16Grok 4.659/s
  17. 17GPT-6.1 Sol58/s
  18. 18Grok 4.754/s

Coding task

All 18 models passed 15 / 15 hidden tests, and all got both trick questions right. Everyday coding is now table stakes; FrontierSWE and CursorBench in the leaderboard are where models still differ.

For 1,000+ models scored on willingness and sensitive knowledge, see the uncensored AI leaderboard (UGI).

05

API pricing

Per million tokens (about 750,000 words). The cheapest model here costs over 154× less than the most expensive.

What a month would cost

messages / month
  1. DeepSeek V4.1 Flash$0.31
  2. Claude Haiku 5.5$0.45
  3. GLM-5.3$0.94
  4. Mistral Large 4$2.31
  5. Gemini 3.8 Flash$3.71
  6. Muse Spark Open$4.42
  7. GPT-6.1 Terra$4.50
  8. Grok 4.6$6.60
  9. Grok 4.7$6.60
  10. GPT-6.1 Sol$9.00
  11. Gemini 4 Argon$9.00
  12. Claude Sonnet 5.5$9.00
  13. Kimi K3.1 Pro$9.00
  14. Kimi K3$9.89
  15. Claude Opus 5.5$18
  16. Muse Wonder$18
  17. GPT-6 Astra$45
  18. Claude Fable 5.1$45

~1,500 tokens in (your message plus chat history) and ~600 out per reply. List prices via the API, no discounts. Reasoning (“thinking”) tokens are billed as output, so heavy-thinking settings can cost several times more. Chat apps with a monthly plan price differently.

Model, cheapest firstInputOutputContext
DeepSeek V4.1 FlashDeepSeek$0.006$0.501.05M
Claude Haiku 5.5Anthropic$0.10$0.501M
GLM-5.3Z.ai$0.070$1.401.05M
Mistral Large 4Mistral AI$0.70$2.101.05M
Gemini 3.8 FlashGoogle$0.82$4.131.05M
GPT-6.1 TerraOpenAI$1$51.05M
Muse Spark OpenMeta$1.25$4.251.05M
Grok 4.6xAI$2$6500K
Grok 4.7xAI$2$6500K
GPT-6.1 SolOpenAI$2$101.05M
Gemini 4 ArgonGoogle$2$101.05M
Claude Sonnet 5.5Anthropic$2$101M
Kimi K3.1 ProMoonshot AI$2$101.05M
Kimi K3Moonshot AI$0.19$161.05M
Claude Opus 5.5Anthropic$4$201M
Muse WonderMeta$4$201.05M
GPT-6 AstraOpenAI$10$501.05M
Claude Fable 5.1Anthropic$10$501M

List prices from OpenRouter. Providers may offer batch or caching discounts.

06

Every test, explained

What each benchmark actually asks, who made it, how to read the score, and the catch.

07

Head to head

Which model wins each test, side-by-side answers and websites, and what each costs.

Popular comparisons

Claude Opus 5.5 vs GPT-6 AstraGPT-6.1 Sol vs Claude Sonnet 5.5Gemini 4 Argon vs GPT-6.1 SolDeepSeek V4.1 Flash vs Kimi K3Grok 4.7 vs Gemini 4 ArgonClaude Haiku 5.5 vs Gemini 3.8 FlashGLM-5.3 vs DeepSeek V4.1 FlashMuse Wonder vs Claude Opus 5.5Claude Fable 5.1 vs GPT-6 AstraMistral Large 4 vs DeepSeek V4.1 Flash

Every model

GPT-6 AstraGPT-6.1 SolClaude Opus 5.5GPT-6.1 TerraClaude Fable 5.1Gemini 4 ArgonClaude Sonnet 5.5Muse WonderClaude Haiku 5.5DeepSeek V4.1 FlashGemini 3.8 FlashKimi K3Grok 4.6GLM-5.3Grok 4.7Muse Spark OpenMistral Large 4Kimi K3.1 Pro
08

Questions

What is an AI benchmark?

A fixed set of tasks with known right answers (or a fair way to judge them) that every AI model is run on, so models can be compared on the same footing. Each benchmark tests one thing: science reasoning, competition maths, building websites, staying factual, running a task on its own.

Which AI model is the best right now?

On Epoch AI’s overall capability index, GPT-6 Astra from OpenAI is currently top of the models we track, followed by GPT-6.1 Sol and Claude Opus 5.5. The gaps at the top are small, and the best model depends on the job — see the category tabs in the leaderboard.

Which AI is best for coding?

Claude Opus 5.5 leads WebDev Arena, where people vote on which model built the better website. For long, real-world coding jobs, see CursorBench and FrontierSWE — the leaders there aren’t always the same.

What is the best open-weights AI model?

DeepSeek V4.1 Flash from DeepSeek has the highest overall score among the open-weights models here. Open weights means you can download the model and run it on your own hardware.

Which AI model is the cheapest?

Of the models here, DeepSeek V4.1 Flash ($0.006 in / $0.50 out per million tokens) and Claude Haiku 5.5 ($0.10 in / $0.50 out) are the cheapest through the API. The most expensive is Claude Fable 5.1 at $10 in / $50 out. Use the cost calculator to see what that means for your usage.

Can AI benchmarks be trusted?

Partly. A model can be trained on test questions that leaked online, some tests are run by the company selling the product, and many older tests are now so easy that top models all score near 100%. That’s why we show many tests side by side, say who made each one, and point out what each score doesn’t tell you.

What is the Epoch Capabilities Index (ECI)?

The Epoch Capabilities Index (ECI) is Epoch AI’s single overall score. It combines dozens of benchmarks into one scale, giving harder tests more weight, so a model can be ranked even when it hasn’t been run on every test. It works like a difficulty scale, not a percentage: higher is better, and a few points is a real gap at the top. Each score has an uncertainty range of about ±2–3 points.

Where do these numbers come from?

Benchmark scores come from Epoch AI’s public benchmarking hub (CC BY 4.0), which runs some tests itself and collects official results for others. Prices and context sizes come from OpenRouter. We didn’t run the tests ourselves; we explain them and put them side by side.

Benchmark scores: Epoch AI, Capabilities & benchmarking, used under CC BY 4.0. For each model we show its best published result across reasoning settings. Prices and context sizes: OpenRouter. Data refreshed November 9, 2026. Hands-on tests (answers, websites, code, refusals, speed) are our own single runs. Explanations written by Cleus.