AI rankings for math and science

CritPt

Often citedHigher is better

Made by 50+ physicists led by Argonne National Laboratory and the University of Illinois: 71 unpublished physics problems at the level of early research projects rather than textbook exercises. Answers are numbers, formulas, or code, graded automatically, and scores are very low. Scores run from 0 to 100%. Higher is better.

Top score32.3%GPT-5.6 Sol
Models tested78
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 🥇OpenAIGPT-5.6 SolReasoning effort: Max49.5%32.3%1494
  2. 🥈AnthropicClaude Opus 5.5Reasoning effort: Max61.4%31.7%1511
  3. 🥈OpenAIGPT-6 AstraReasoning effort: Max54.7%31.7%1490
  4. 🥈OpenAIGPT-6.1 SolReasoning effort: Max52.9%31.7%—
  5. #5AnthropicClaude Sonnet 5.5Reasoning effort: Max55.0%31.4%—
  6. #6AnthropicClaude Fable 5.1Reasoning effort: Extra High59.1%31.1%1518
  7. #7OpenAIGPT-6 SolReasoning effort: Max47.9%30.9%1451
  8. #8OpenAIGPT-5.6 TerraReasoning effort: Max42.9%30.0%1479
  9. #8OpenAIGPT-5.4 ProReasoning effort: Extra High—30.0%—
  10. #10AnthropicClaude Opus 5Reasoning effort: Max54.9%29.1%1516
  11. #11AnthropicClaude Fable 5Reasoning effort: Max55.5%28.6%1522
  12. #12GoogleGemini 4 ArgonReasoning effort: High57.1%27.1%1530
  13. #12OpenAIGPT-5.5Reasoning effort: Extra High45.8%27.1%1500
  14. #14XiaomiMiMo-V2.6-Pro49.4%26.6%1487
  15. #15MetaMuse Spark 1.3Reasoning effort: Extra High48.7%26.0%1509
  16. #16OpenAIGPT-5.4Reasoning effort: Extra High43.7%23.4%1492
  17. #17Moonshot AIKimi K3Reasoning effort: Max46.9%23.4%1501
  18. #18AnthropicClaude Opus 4.8Reasoning effort: Max48.7%20.9%1477
  19. #18Z.aiGLM 5.2Reasoning effort: Max41.1%20.9%1488
  20. #20OpenAIGPT-5.6 LunaReasoning effort: Max39.5%20.6%1477
  21. #21xAIGrok 4.6Reasoning effort: Extra High44.1%19.7%1448
  22. #22OpenAIGPT-6 LunaReasoning effort: Max38.5%19.4%1460
  23. #23Z.aiGLM 5.3Reasoning effort: Max42.3%19.1%1494
  24. #24GoogleGemini 3.8 FlashReasoning effort: High47.8%18.3%1518
  25. #25xAIGrok 4.7Reasoning effort: High43.1%18.0%1444
  26. #25DeepSeekDeepSeek V4 ProReasoning effort: Max41.0%18.0%1445
  27. #27GoogleGemini 3.1 Pro47.0%17.7%1488
  28. #27MetaMuse Spark 1.2Reasoning effort: Extra High45.5%17.7%1468
  29. #29AnthropicClaude Sonnet 5Reasoning effort: Max41.3%16.9%1474
  30. #30DeepSeekDeepSeek V4 FlashReasoning effort: Max38.6%16.6%1425
  31. #31xAIGrok 4.5Reasoning effort: High42.7%15.4%1473
  32. #31Z.aiGLM 5.3 Flash39.9%15.4%1503
  33. #33MetaMuse Spark 1.1Reasoning effort: Extra High46.2%15.1%1492
  34. #34GoogleGemini 3.7 FlashReasoning effort: High47.9%14.3%1503
  35. #34DeepSeekDeepSeek V4.1 FlashReasoning effort: Max39.2%14.3%1503
  36. #36AlibabaQwen3.7 Max40.5%13.4%1484
  37. #37GoogleGemini 3.5 FlashReasoning effort: High42.7%13.1%1482
  38. #38OpenAIGPT-5Reasoning effort: High28.5%12.6%1434
  39. #39AnthropicClaude Opus 4.7Reasoning effort: Max42.3%12.0%1492
  40. #39XiaomiMiMo-V2.6-Flash35.1%12.0%1472
  41. #41MetaMuse Spark40.7%11.3%1466
  42. #42GoogleGemini 3.6 FlashReasoning effort: High40.8%10.6%1501
  43. #43Moonshot AIKimi K2.7 Code35.0%10.0%—
  44. #43OpenAIGPT-5.4 MiniReasoning effort: Extra High28.1%10.0%1439
  45. #45OpenAIGPT-5.4 NanoReasoning effort: Extra High28.3%9.3%1425
  46. #46AlibabaQwen3.7 Plus35.6%9.1%1466
  47. #46xAIGrok Build 0.1—9.1%—
  48. #48Moonshot AIKimi K2.637.5%8.0%1480
  49. #48xAIGrok 4.3Reasoning effort: High37.2%8.0%1419
  50. #50Z.aiGLM-5.130.1%4.6%1475
  51. #51XiaomiMiMo V2.5 Pro35.7%4.0%1481
  52. #52MiniMaxMiniMax M339.0%3.7%1432
  53. #52XiaomiMiMo V2.527.2%3.7%1438
  54. #52Ring AIRing-2.6-1T21.6%3.7%—
  55. #55AnthropicClaude Sonnet 4.6Reasoning effort: Max33.6%3.1%1464
  56. #55Moonshot AIKimi K2.530.7%3.1%1471
  57. #55NVIDIANemotron 3 Ultra28.4%3.1%1442
  58. #55NVIDIANemotron 3 Super20.8%3.1%1373
  59. #59AlibabaQwen3.6 Plus27.8%2.9%1454
  60. #60StepfunStep 3.7 Flash21.4%2.3%—
  61. #61GoogleGemini 2.5 Pro22.5%2.0%1439
  62. #62GoogleGemma 4 31B23.6%1.4%1469
  63. #63OpenAIGPT OSS 120BReasoning effort: High19.6%1.1%1381
  64. #63AnthropicClaude Sonnet 4.517.8%1.1%1428
  65. #63GoogleGemini 3.1 Flash Lite17.2%1.1%1436
  66. #66GoogleGemini 2.5 Flash12.1%1.1%1405
  67. #67Arcee AITrinity Large Thinking15.8%0.9%1384
  68. #68MiniMaxMiniMax M2.729.6%0.6%1422
  69. #69AnthropicClaude Opus 412.3%0.3%1421
  70. #70AlibabaQwen3.6 35B A3B22.2%0.3%—
  71. #70AnthropicClaude Sonnet 410.7%0.3%1404
  72. #72OpenAIGPT-5 MiniReasoning effort: High21.5%0.0%1405
  73. #72GoogleGemini 3.5 Flash-Lite18.8%0.0%1442
  74. #72Mistral AIMistral Medium 3.513.8%0.0%1429
  75. #72AnthropicClaude Haiku 4.510.4%0.0%1400
  76. #72UpstageSolar Pro 310.3%0.0%—
  77. #72MetaLlama 4 Maverick4.9%0.0%1316
  78. #72MetaLlama 4 Scout3.8%0.0%1309
Hard test · top score 32.3% · Half of models ≤ 12.0%Last updated 2026-10-07

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.