Benchmark Almanac

Build your own index

Every intelligence index is a weighted average of benchmark scores, and somebody picked the weights. Here, you pick them. See who comes out on top, and whether the ranking holds when the weights shift.

Start from

The benchmarks the most labs reported in 2026 (33 labs down to 20), minus any where top models already score past 92%. We use the version most recent models were scored on.

Tests and weights

Pick from 1,201 benchmarks with at least 8 scored models.

  1. Terminal-Bench 2.164 models13%
  2. SWE-Bench Pro97 models13%
  3. Humanity's Last Exam181 models13%
  4. LiveCodeBench v6110 models13%
  5. DeepSWE69 models13%
  6. MMLU-Pro295 models13%
  7. MMMU-Pro131 models13%
  8. AutomationBench72 models13%

Scores from

We use independent runs where they exist and fill the gaps with labs' own numbers.

A model needs scores on

Where a model skipped a test, we predict its score from its other results and how hard that test was.

Who comes out on top?

Kimi K3.1 Pro leads this mix with 82.9 and stays first under 80% of nearby weightings. Claude Fable 5.2 is second at 81.9. We rank the 85 models with scores on at least 4 of the 8 tests, from independent runs and labs’ own numbers.

  1. 82.9Kimi K3.1 Pro
  2. 81.9Claude Fable 5.2
  3. 79.3Claude Opus 5.5
  4. 77.5GLM-5.3
  5. 77.5LongCat-2.5
  6. 76.0GPT-6 Astra
  7. 75.1Claude Opus 5
  8. 75.1GLM-5.3-Flash
  9. 74.9Claude Sonnet 5.5
  10. 74.7Inkling 2 Small
  11. 74.0Step 5 Preview
  12. 73.7DeepSeek V4 Pro 0813
  13. 73.1Claude Fable 5
  14. 73.1Gemini 3.8 Flash
  15. 72.9GPT-5.6 Sol
  16. 72.0DeepSeek V4.1 Flash
  17. 71.9Kimi K3
  18. 71.5Tencent Hy4 preview
  19. 71.2Claude Fable 5.1
  20. 71.1Ling-3.1-flash
  21. 70.2Qwen 3.8 Max
  22. 68.7Claude Opus 4.8
  23. 68.4GPT-5.6 Terra
  24. 68.3Beam 1.0
  25. 67.9Muse Spark 1.1
  26. 67.0GPT-5.5
  27. 65.8GPT-5.6 Luna
  28. 65.8DeepSeek-V4-Pro
  29. 65.7Claude Sonnet 5
  30. 65.4Qwen3.8-Omni-Flash
  31. 65.1Qwen3.8 Flash
  32. 64.5GLM-5.2
  33. 63.2Gemini 3.6 Flash
  34. 61.9Seed2.1 Pro
  35. 61.1DeepSeek-V4-Flash
  36. 61.0Claude Opus 4.7
  37. 61.0Qwen 3.8 27B
  38. 61.0Tencent Hy3
  39. 60.4Gemini 3.5 Flash
  40. 60.4Intern-S2-397B
  1. 1Kimi K3.1 Pro82.9
  2. 2Claude Fable 5.281.9
  3. 3Claude Opus 5.579.3
  4. 4GLM-5.377.5
  5. 5LongCat-2.577.5
  6. 6GPT-6 Astra76.0
  7. 7Claude Opus 575.1
  8. 8GLM-5.3-Flash75.1
  9. 9Claude Sonnet 5.574.9
  10. 10Inkling 2 Small74.7
  11. 11Step 5 Preview74.0
  12. 12DeepSeek V4 Pro 081373.7
  13. 13Claude Fable 573.1
  14. 14Gemini 3.8 Flash73.1
  15. 15GPT-5.6 Sol72.9

Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards

RankModelIndexTestsBestRank rangeWorstFirst
1Kimi K3.1 Pro82.94 of 81280%
2Claude Fable 5.281.94 of 81221%
3Claude Opus 5.579.34 of 833
4GLM-5.377.54 of 845
5LongCat-2.577.54 of 846
6GPT-6 Astra76.04 of 857
7Claude Opus 575.14 of 8512
8GLM-5.3-Flash75.14 of 8615
9Claude Sonnet 5.574.94 of 8715
10Inkling 2 Small74.74 of 8716
11Step 5 Preview74.05 of 8817
12DeepSeek V4 Pro 081373.74 of 8816
13Claude Fable 573.15 of 8819
14Gemini 3.8 Flash73.14 of 8918
15GPT-5.6 Sol72.95 of 8818
16DeepSeek V4.1 Flash72.06 of 8631
17Kimi K371.95 of 81420
18Tencent Hy4 preview71.55 of 81520
19Claude Fable 5.171.24 of 81522
20Ling-3.1-flash71.15 of 81223
21Qwen 3.8 Max70.26 of 81822
22Claude Opus 4.868.75 of 82025
23GPT-5.6 Terra68.45 of 82026
24Beam 1.068.35 of 82129
25Muse Spark 1.167.94 of 82127
26GPT-5.567.05 of 82328
27GPT-5.6 Luna65.85 of 82532
28DeepSeek-V4-Pro65.85 of 82633
29Claude Sonnet 565.75 of 82531
30Qwen3.8-Omni-Flash65.44 of 82532
31Qwen3.8 Flash65.14 of 82833
32GLM-5.264.55 of 82734
33Gemini 3.6 Flash63.24 of 83236
34Seed2.1 Pro61.95 of 83140
35DeepSeek-V4-Flash61.16 of 83349
36Claude Opus 4.761.04 of 83444
37Qwen 3.8 27B61.05 of 83448
38Tencent Hy361.04 of 83447
39Gemini 3.5 Flash60.46 of 83650
40Intern-S2-397B60.44 of 83451
41GPT-5.460.36 of 83548
42Kimi K2.659.85 of 83453
43Seed2.0 Pro59.65 of 83451
44Inkling-Small59.54 of 83648
45Qwen3.5 397B-A17B59.24 of 83650
46GLM-5.158.74 of 83653
47Inkling58.64 of 84051
48Qwen 3.6 Plus58.46 of 84153
49Gemini 3 Pro58.25 of 84053
50Nemotron 3.5 Super VL57.76 of 83959

How much do the weights matter?

In the default mix, 4 of the top 10 models can move three places or more when the weights shift.

How it works

We put every test on a 0 to 100 scale. Percentages stay as they are. Elo and other units run from the lowest model to the highest, and tests where lower is better get flipped. A model's index is the weighted average.

Where a model skipped a test, we predict that score from its other results and the test's difficulty, so skipping a hard test doesn't help and taking one doesn't hurt. Hover a model to see which of its scores are predicted.

Why weights matter

Two indexes built from the same scores can disagree on who's first, because the weights decide which tests count most. Published indexes change their weights between versions, and every change reorders some models.

So we rerank every model under 400 weightings near yours, each weight moved up or down by as much as half. Rank range shows each model's best and worst place, and First is the share of those weightings it wins.

What it leaves out

Independent runs cover far fewer new models than labs' own posts, so the board starts with both. Switch to Independent runs to leave lab numbers out, though the newest models may drop off.

Predictions assume a model's strengths carry across tests, so they miss a model that's strong at one thing and weak at another. Raise the minimum to lean on them less. And no average says how sure to be about any one score.