Best AI for reasoning 2026.

Today's AI models, ranked by how well they reason on two tests from an independent evaluator. One with expert-⁠level questions across many subjects and one with physics research problems.

Data from 13 November 2026.

  • The best at HLE

    Claude Fable 5.2

    80.2% on HLE

  • The best at CritPt

    Claude Fable 5.2

    76.5% on CritPt

  • The best open-⁠weights model

    Kimi K3.1 Pro

    37.3% on HLE

  • The cheapest per task

    GPT-6.1 Terra

    $0.96 per task, among the top ten

Sort by

58 models

#Model
1Claude Fable 5.2 Anthropic · max effort80.2%76.5%65$26.54$8 and $409 Nov 2026
2Gemini 4 Carbon Google · high effort69.1%69.0%60$5.48$2 and $10 $4 and $20 from 1 December10 Nov 2026
3Claude Opus 5.5 Anthropic · max effort67.9%68.0%59$10.04$4 and $2022 Sept 2026
4GPT-6 Astra OpenAI · max effort63.8%70.5%57$6.24$10 and $503 Sept 2026
5GPT-6.1 Sol OpenAI · max effort58.8%69.5%55$1.27$2 and $1029 Sept 2026
6Gemini 4 Hydrogen Google · extra high effort58.8%66.0%54$3.16$1.50 and $7.5012 Nov 2026
7Gemini 4 Argon Google · high effort56.7%52.5%50$4.97$2 and $10 $4 and $20 from 1 December13 Oct 2026
8Claude Sonnet 5.5 Anthropic · max effort53.3%59.0%55$10.02$2 and $1028 Sept 2026
9Muse Wonder Full Meta · max effort48.5%56.5%50No dataNo per-token price12 Oct 2026
10GPT-6.1 Terra OpenAI · max effort47.3%68.0%51$0.96$1 and $526 Oct 2026
11Muse Wonder Meta · max effort39.0%50.5%44$6.95$4 and $2012 Oct 2026
12Kimi K3.1 Pro Moonshot AI · open weights37.3%49.5%46$3.68$2 and $1024 Oct 2026
13Gemini 3.8 Flash Google · high effort33.1%46.5%31$7.77$0.75 and $3.75 $1.50 and $7.50 from 1 January 20272 Sept 2026
14Tencent Hy4 Tencent32.7%39.0%35$1.03$0.834 and $2.501 half price until 31 December4 Nov 2026
15Claude Haiku 5.5 Anthropic · max effort29.6%43.5%36$0.41$0.10 and $0.507 Oct 2026
16Inkling 2 Small Thinking Machines · open weights28.7%35.0%31$1.18$0.30 and $15 Nov 2026
17MiniMax-M3.1-Flash MiniMax · open weights28.1%32.5%29$0.49$0.15 and $0.6019 Oct 2026
18Muse Spark Open Meta · extra high effort · open weights26.3%29.5%25$2.02No per-token price15 Oct 2026
19DeepSeek V4.1 Flash 1017 DeepSeek · max effort · open weights25.3%25.0%33$0.60$0.30 and $1.2017 Oct 2026
20Muse Ember Meta · max effort25.1%39.0%36$0.14$0.10 and $0.509 Nov 2026
  1. 1Claude Fable 5.2 Anthropic · max effort80.2%
    HLE Diamond
    80.2%
    CritPt v2, hard subset
    76.5%
    Score
    65
    Cost per task
    $26.54
    Price per million tokens
    $8 and $40
    Released
    9 Nov 2026
  2. 2Gemini 4 Carbon Google · high effort69.1%
    HLE Diamond
    69.1%
    CritPt v2, hard subset
    69.0%
    Score
    60
    Cost per task
    $5.48
    Price per million tokens
    $2 and $10 $4 and $20 from 1 December
    Released
    10 Nov 2026
  3. 3Claude Opus 5.5 Anthropic · max effort67.9%
    HLE Diamond
    67.9%
    CritPt v2, hard subset
    68.0%
    Score
    59
    Cost per task
    $10.04
    Price per million tokens
    $4 and $20
    Released
    22 Sept 2026
  4. 4GPT-6 Astra OpenAI · max effort63.8%
    HLE Diamond
    63.8%
    CritPt v2, hard subset
    70.5%
    Score
    57
    Cost per task
    $6.24
    Price per million tokens
    $10 and $50
    Released
    3 Sept 2026
  5. 5GPT-6.1 Sol OpenAI · max effort58.8%
    HLE Diamond
    58.8%
    CritPt v2, hard subset
    69.5%
    Score
    55
    Cost per task
    $1.27
    Price per million tokens
    $2 and $10
    Released
    29 Sept 2026
  6. 6Gemini 4 Hydrogen Google · extra high effort58.8%
    HLE Diamond
    58.8%
    CritPt v2, hard subset
    66.0%
    Score
    54
    Cost per task
    $3.16
    Price per million tokens
    $1.50 and $7.50
    Released
    12 Nov 2026
  7. 7Gemini 4 Argon Google · high effort56.7%
    HLE Diamond
    56.7%
    CritPt v2, hard subset
    52.5%
    Score
    50
    Cost per task
    $4.97
    Price per million tokens
    $2 and $10 $4 and $20 from 1 December
    Released
    13 Oct 2026
  8. 8Claude Sonnet 5.5 Anthropic · max effort53.3%
    HLE Diamond
    53.3%
    CritPt v2, hard subset
    59.0%
    Score
    55
    Cost per task
    $10.02
    Price per million tokens
    $2 and $10
    Released
    28 Sept 2026
  9. 9Muse Wonder Full Meta · max effort48.5%
    HLE Diamond
    48.5%
    CritPt v2, hard subset
    56.5%
    Score
    50
    Cost per task
    No data
    Price per million tokens
    No per-token price
    Released
    12 Oct 2026
  10. 10GPT-6.1 Terra OpenAI · max effort47.3%
    HLE Diamond
    47.3%
    CritPt v2, hard subset
    68.0%
    Score
    51
    Cost per task
    $0.96
    Price per million tokens
    $1 and $5
    Released
    26 Oct 2026
  11. 11Muse Wonder Meta · max effort39.0%
    HLE Diamond
    39.0%
    CritPt v2, hard subset
    50.5%
    Score
    44
    Cost per task
    $6.95
    Price per million tokens
    $4 and $20
    Released
    12 Oct 2026
  12. 12Kimi K3.1 Pro Moonshot AI · open weights37.3%
    HLE Diamond
    37.3%
    CritPt v2, hard subset
    49.5%
    Score
    46
    Cost per task
    $3.68
    Price per million tokens
    $2 and $10
    Released
    24 Oct 2026
  13. 13Gemini 3.8 Flash Google · high effort33.1%
    HLE Diamond
    33.1%
    CritPt v2, hard subset
    46.5%
    Score
    31
    Cost per task
    $7.77
    Price per million tokens
    $0.75 and $3.75 $1.50 and $7.50 from 1 January 2027
    Released
    2 Sept 2026
  14. 14Tencent Hy4 Tencent32.7%
    HLE Diamond
    32.7%
    CritPt v2, hard subset
    39.0%
    Score
    35
    Cost per task
    $1.03
    Price per million tokens
    $0.834 and $2.501 half price until 31 December
    Released
    4 Nov 2026
  15. 15Claude Haiku 5.5 Anthropic · max effort29.6%
    HLE Diamond
    29.6%
    CritPt v2, hard subset
    43.5%
    Score
    36
    Cost per task
    $0.41
    Price per million tokens
    $0.10 and $0.50
    Released
    7 Oct 2026
  16. 16Inkling 2 Small Thinking Machines · open weights28.7%
    HLE Diamond
    28.7%
    CritPt v2, hard subset
    35.0%
    Score
    31
    Cost per task
    $1.18
    Price per million tokens
    $0.30 and $1
    Released
    5 Nov 2026
  17. 17MiniMax-M3.1-Flash MiniMax · open weights28.1%
    HLE Diamond
    28.1%
    CritPt v2, hard subset
    32.5%
    Score
    29
    Cost per task
    $0.49
    Price per million tokens
    $0.15 and $0.60
    Released
    19 Oct 2026
  18. 18Muse Spark Open Meta · extra high effort · open weights26.3%
    HLE Diamond
    26.3%
    CritPt v2, hard subset
    29.5%
    Score
    25
    Cost per task
    $2.02
    Price per million tokens
    No per-token price
    Released
    15 Oct 2026
  19. 19DeepSeek V4.1 Flash 1017 DeepSeek · max effort · open weights25.3%
    HLE Diamond
    25.3%
    CritPt v2, hard subset
    25.0%
    Score
    33
    Cost per task
    $0.60
    Price per million tokens
    $0.30 and $1.20
    Released
    17 Oct 2026
  20. 20Muse Ember Meta · max effort25.1%
    HLE Diamond
    25.1%
    CritPt v2, hard subset
    39.0%
    Score
    36
    Cost per task
    $0.14
    Price per million tokens
    $0.10 and $0.50
    Released
    9 Nov 2026

HLE Diamond and CritPt v2's hard subset, the score and the cost per task from Artificial Analysis, read on 13 November 2026 with each model at its highest effort. The price and the release date from each maker's page. HLE Diamond and CritPt v2 on Artificial Analysis.

Questions.

  • Which one for your company's software?

    No single one wins at everything. We explain it in the article “Which AI Model to Use to Build Your Company's Software”, including how we split the work between them.

  • What is the best AI for reasoning?

    By Artificial Analysis's tests on 13 November 2026, Claude Fable 5.2, from Anthropic, which wins at both, with 80.2% of HLE Diamond and 76.5% of CritPt's hard subset.

  • What does HLE Diamond measure?

    Whether the model can answer 1,000 expert-⁠level questions in mathematics, science and the humanities, half of them about reasoning and half about knowledge. It is a revised version of Humanity's Last Exam, from the Center for AI Safety and Scale AI.

  • What does CritPt measure?

    Whether the model can solve physics research problems with a computer to run code and no internet. It comes from Argonne and the University of Illinois, and we show its hardest part, which Artificial Analysis publishes on its own.

  • What is the cost per task?

    What it costs on average for the model to do one of the tasks behind Artificial Analysis's score, across all its tests. It says more than the price per million tokens, because for the same work some models use many more tokens than others.

  • And the whole comparison?

    In “AI comparison 2026”, with every model ranked by its score, what each task costs and how long it takes, its price and its context.

Let's talk.

Tell us what you need and one of the partners who write the code gets back to you.