Benchmark Almanac

How do new models get tested?

Labs named a median of 29 benchmarks for each frontier model of the past year, and others scored the same models on a median of 387. Only 14% of what labs name has an independent run of the same model.

How much of a launch post can anyone check?

Of the benchmarks labs named in launch posts since 2024, an official leaderboard or Epoch AI has scored the same model on 14%, about the same share each year. For another 63%, someone scores the benchmark but hasn't run that model. OpenAI's posts are the easiest to check, at 32%, and Qwen's the hardest of the frontier labs, at 9%.

How do claims and official scores differ?
  1. OpenAI 32% 39% 29% 32%746
  2. xAI 30% 43% 27% 30%152
  3. Moonshot AI 26% 54% 20% 26%300
  4. Anthropic 25% 40% 35% 25%705
  5. Google DeepMind 21% 48% 31% 21%778
  6. DeepSeek 19% 59% 22% 19%362
  7. Z.ai 18% 67% 15% 18%348
  8. Meta 16% 53% 31% 16%316
  9. Qwen 71% 20% 9%1,645
  10. Other labs 74% 20% 6%4,060

Sources: Labs’ launch posts, read by Benchmark Almanac, official leaderboards and Epoch AI (CC BY 4.0)

What do others test beyond the launch post?

The 38 frontier models released since November 21st, 2025 have independent scores on a median of 387 benchmarks, about 13 times the 29 in their launch posts. The two lists share a median of ten. Gemini 3.1 Pro has the most independent scores, 719.

  1. GPT-6.2 Astra 29137
  2. Claude Fable 5.2 33210
  3. Gemini 4 Carbon 16210
  4. GPT-6.1 Terra 9210
  5. Gemini 4 Argon 27445
  6. GPT-6.1 Sol 9522
  7. Claude Sonnet 5.5 36449
  8. Claude Opus 5.5 38648
  9. GPT-6 Sol 9334
  10. GPT-6 Astra 33686
  11. Gemini 3.8 Flash 13547
  12. Claude Fable 5.1 37446
  13. Gemini 3.7 Flash 18298
  14. Claude Opus 5 47554
  15. Kimi K3 66556
  16. GPT-5.6 Sol 33643
  17. GPT-5.6 Terra 32356
  18. Claude Sonnet 5 40405
  19. Claude Fable 5 30336
  20. Claude Opus 4.8 48513
  21. Qwen3.7-Max 40163
  22. Gemini 3.5 Flash 13349
  23. GPT-5.5 28594
  24. GPT-5.5 Pro 835
  25. Claude Opus 4.7 48385
  26. Muse Spark 2947
  27. Grok 4.20 4184
  28. GPT-5.4 29491
  29. GPT-5.4 Pro 1132
  30. Gemini 3.1 Pro 16719
  31. Claude Sonnet 4.6 40427
  32. Claude Opus 4.6 42463
  33. Kimi K2.5 48305
  34. Gemini 3 Flash 21388
  35. GPT-5.2 19417
  36. GPT-5.2 Pro 937
  37. DeepSeek-V3.2 23269
  38. Claude Opus 4.5 32325

Sources: Labs’ launch posts, read by Benchmark Almanac, official leaderboards and Epoch AI (CC BY 4.0)

How long does a model stay at the frontier?

Since 2024, a model that came in at number one on Epoch AI's Capabilities Index stayed in the top ten for a median of 203 days, about seven months. One that came in second to fifth stayed 117 days, and one that came in lower, 35.

0% 25% 50% 75% 100% 0 1 2 3 4 5 6 7 8 Months since release Came in at number one Came in sixth to tenth Came in second to fifth 0% 25% 50% 75% 100% 0 1 2 3 4 5 6 7 8 Months since release

Epoch Capabilities Index, Epoch AI (CC BY 4.0)

How this is counted

A launch post counts with the system cards, model cards and technical reports published with it, as on What labs report. A benchmark counts once per release, and versions stay apart, so SWE-bench Verified and SWE-bench Pro are two benchmarks. Safety, alignment, bio and cyber evaluations are left out, since few labs publish them and few boards run them.

An independent score is one from a benchmark's official leaderboard, from Epoch AI's own runs, or from another board Epoch AI copied. It has to be the model itself. An agent someone else built around the model doesn't count. We read 4,766 leaderboards but not every one, and from Artificial Analysis we read only its index numbers. A benchmark that only boards we don't read have run counts as having no independent score.

The frontier labs are the nine that the reporting page follows. Other labs are grouped. A frontier model is one in the top ten on Epoch AI's Capabilities Index among the models out the day its launch post came out.

A model stays in the top ten until ten models released after it score higher on the index, counted by release date, so a model scored later still counts from the day it came out. Models still in the top ten haven't left yet, so each line is a survival estimate, and a line stops when fewer than five models are left to follow. Of the nine models that came in at number one in 2026, five are still in the top ten.

More insights