Benchmark Almanac

Progress over time

Pick up to eight benchmarks to see the best score on each, by the release date of the model that set it.

Model progress over time

In the past year, Humanity's Last Exam climbed the most, from 31.6% to 54.8%. GPT-6 Astra holds the best score now.

  • Humanity's Last Exam (Center for AI Safety and Scale AI)
  • SWE-bench Verified (Princeton NLP and OpenAI)
  • SWE-Bench Pro (Scale AI)
  • MMMU-Pro (MMMU team)
0102030405060708090100Nov 2025DecJan 2026FebMarAprMayJunJulAugSepOctNovHumanity's Last ExamGPT-6 Astra 54.8%SWE-bench VerifiedClaude Opus 4.7 83.5%SWE-Bench ProMuse Spark 1.1 61.5%MMMU-ProGPT-5.6 Sol 84.6%020406080100Jan 2026AprJulOct

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

BenchmarkAt the startNowChangeHolds the best scoreNew recordsScores from
Humanity's Last Exam31.6%54.8%+23.2 ptsGPT-6 Astra4Official leaderboard
SWE-bench Verified73.5%83.5%+9.9 ptsClaude Opus 4.73Epoch AI runs
SWE-Bench Pro43.6%61.5%+17.9 ptsMuse Spark 1.14Official leaderboard
MMMU-Pro78.4%84.6%+6.2 ptsGPT-5.6 Sol6Labs’ own numbers

Filled dots are independent runs. Rings are numbers a lab reports about its own model. A line ends at the newest model its source has scored.