HLE (Humanity's Last Exam)
- What does it measure?
- The hardest exam there is, built from questions at the frontier of human knowledge. It gathers problems from over a hundred fields — mathematics, humanities, natural sciences, engineering — that are difficult for anyone but a specialist in that field.
- Who evaluates it, and how?
- The Center for AI Safety and Scale AI collected 2,500 questions from roughly 1,000 experts at 500+ institutions in 50 countries. About 76% are short-answer questions requiring an exact answer and 24% are multiple choice. Artificial Analysis scores the 2,158 text-only questions, excluding those that need an image.
- Reading the score
- 0–100%. As of early 2026 even the best models sit around 40%. A low score does not mean 'bad' — the exam is extreme — and a few percentage points is a large difference in ability. A private held-out set exists alongside the public questions to detect memorisation.
Rank
Top score61.4%
Models tested122
Scoring basisHigher is better
Last updated2026.11.21
Model release date
Can it solve research-level physics problems?CritPt
Gemini 4 Carbon🥇
Gemini 4 Carbon🥇CritPt
Gemini 4 Carbon🥇Which AI gives the math answers people prefer most?Arena Math
Gemini 4 Argon🥇
Gemini 4 Argon🥇Arena Math
Gemini 4 Argon🥇
Can it solve problems that stump experts across fields?
Humanity's Last Exam
- 🥇AnthropicClaude Opus 5.561.4%31.7%1517
- 🥈AnthropicClaude Fable 5.159.1%31.1%1530
- 🥉GoogleGemini 4 Argon57.1%27.1%1535
- #4AnthropicClaude Fable 555.5%28.6%1521
- #5AnthropicClaude Sonnet 5.555.0%31.4%1497
- #6AnthropicClaude Opus 554.9%29.1%1516
- #7OpenAIGPT-6 Astra54.7%31.7%1477
- #8OpenAIGPT-6.1 Sol52.9%31.7%1499
- #9Moonshot AIKimi K3.1 Pro52.5%24.6%1477
- #10MetaMuse Wonder50.1%—1504
- #11OpenAIGPT-5.6 Sol49.5%32.3%1494
- #12XiaomiMiMo-V2.6-Pro49.4%26.6%1482
- #13AnthropicClaude Opus 4.848.7%20.9%1475
- #13MetaMuse Spark 1.348.7%26.0%1502
- #15MiniMaxMiniMax M3-1 Flash48.0%——
- #16GoogleGemini 3.7 Flash47.9%14.3%1501
- #16OpenAIGPT-6 Sol47.9%30.9%1444
- #18GoogleGemini 3.8 Flash47.8%18.3%1519
- #19GoogleGemini 3.1 Pro47.0%17.7%1488
- #20Moonshot AIKimi K346.9%23.4%1502
- #21StepfunStep-546.5%20.9%1474
- #22OpenAIGPT-6.1 Terra46.2%28.0%1479
- #22MetaMuse Spark 1.146.2%15.1%1494
- #24OpenAIGPT-5.545.8%27.1%1500
- #25MetaMuse Spark 1.245.5%17.7%1483
- #26AnthropicClaude Haiku 5.544.4%22.6%1447
- #27SpaceXSIGrok 4.644.1%19.7%1445
- #28OpenAIGPT-5.443.7%23.4%1491
- #29DeepSeekDeepSeek V4.1 Flash 101743.3%24.0%—
- #30SpaceXSIGrok 4.743.1%18.0%1443
- #31OpenAIGPT-5.6 Terra42.9%30.0%1478
- #32GoogleGemini 3.5 Flash42.7%13.1%1482
- #32SpaceXSIGrok 4.542.7%15.4%1471
- #34AnthropicClaude Opus 4.742.3%12.0%1490
- #34Z.aiGLM 5.342.3%19.1%1504
- #36AnthropicClaude Sonnet 541.3%16.9%1472
- #37Z.aiGLM 5.241.1%20.9%1488
- #38DeepSeekDeepSeek V4 Pro41.0%18.6%1443
- #39GoogleGemini 3.6 Flash40.8%10.6%1501
- #40MetaMuse Spark40.7%11.3%1468
- #41AlibabaQwen3.7 Max40.5%13.4%1481
- #42AnthropicClaude Opus 4.639.9%—1504
- #42Z.aiGLM 5.3 Flash39.9%15.4%1509
- #44OpenAIGPT-5.6 Luna39.5%20.6%1476
- #45DeepSeekDeepSeek V4.1 Flash39.2%14.3%1496
- #46MiniMaxMiniMax M339.0%3.7%1423
- #47DeepSeekDeepSeek V4 Flash38.6%16.6%1424
- #48OpenAIGPT-6 Luna38.5%19.4%1449
- #49AlibabaQwen3.8 Flash38.0%——
- #50Moonshot AIKimi K2.637.5%8.0%1479
- #51SpaceXSIGrok 4.337.2%8.0%1419
- #52Motif TechnologiesMotif 337.0%——
- #53GoogleGemini 3 Flash36.6%—1475
- #54XiaomiMiMo V2.5 Pro35.7%4.0%1479
- #55AlibabaQwen3.7 Plus35.6%9.1%1465
- #56XiaomiMiMo-V2.6-Flash35.1%12.0%1463
- #57Moonshot AIKimi K2.7 Code35.0%10.0%—
- #57Mistral AIMistral Large 435.0%11.1%1448
- #59SpaceXSIGrok 4.20 (Reasoning)34.5%—1464
- #60MeituanLongCat 2.033.7%——
- #61AnthropicClaude Sonnet 4.633.6%3.1%1461
- #62TencentHy333.5%—1429
- #63AlibabaQwen3.6 Max30.8%—1472
- #64Moonshot AIKimi K2.530.7%3.1%1470
- #65XiaomiMiMo V2 Pro30.4%—1450
- #66AnthropicClaude Opus 4.530.1%—1463
- #66Z.aiGLM-5.130.1%4.6%1474
- #68UpstageSolar Pro 429.8%—1416
- #69SK TelecomA.X K229.6%——
- #69MiniMaxMiniMax M2.729.6%0.6%1423
- #71Z.aiGLM-529.3%—1443
- #72AlibabaQwen3.5 397B A17B29.0%0.9%1451
- #73OpenAIGPT-528.5%12.6%1435
- #74NVIDIANemotron 3 Ultra28.4%3.1%1442
- #75OpenAIGPT-5.4 Nano28.3%9.3%1424
- #76OpenAIGPT-5.4 Mini28.1%10.0%1439
- #77SpaceXSIGrok 4.2027.9%—1449
- #78Z.aiGLM-5-Turbo27.8%——
- #78AlibabaQwen3.6 Plus27.8%2.9%1451
- #80XiaomiMiMo V2.527.2%3.7%1437
- #81UpstageSolar Mini 425.8%——
- #82DeepSeekDeepSeek V3.224.6%—1426
- #83GoogleGemma 4 31B23.6%1.4%1466
- #84GoogleGemini 2.5 Pro22.5%2.0%1439
- #85AlibabaQwen3.6 35B A3B22.2%0.3%—
- #86InclusionAIRing-2.6-1T21.6%3.7%—
- #87OpenAIGPT-5 Mini21.5%0.0%1407
- #88StepfunStep 3.7 Flash21.4%2.3%—
- #89NVIDIANemotron 3 Super20.8%3.1%1370
- #90MiniMaxMiniMax M2.520.5%—1392
- #91OpenAIGPT OSS 120B19.6%1.1%1381
- #92SpaceXSIGrok 4.1 Fast (Reasoning)19.3%—1420
- #93GoogleGemini 3.5 Flash-Lite18.8%0.0%1442
- #94LG AI ResearchK-EXAONE 2.018.6%——
- #95AnthropicClaude Sonnet 4.517.8%1.1%1426
- #96GoogleGemini 3.1 Flash Lite17.2%1.1%1436
- #97Z.aiGLM 5V Turbo17.1%——
- #98Arcee AITrinity Large Thinking15.8%0.9%1383
- #99AlibabaQwen3.5 Plus14.9%——
- #100LG AI ResearchK-EXAONE13.9%——
- #101Mistral AIMistral Medium 3.513.8%0.0%1427
- #102BaiduERNIE 5.0 Thinking13.3%—1434
- #103AnthropicClaude Opus 4.112.5%—1444
- #104AnthropicClaude Opus 412.3%0.3%1421
- #105GoogleGemini 2.5 Flash12.1%1.1%1405
- #106AmazonNova 2 Lite11.6%—1329
- #107NVIDIANemotron 3 Nano 30B A3B11.4%—1347
- #108CohereNorth Mini Code11.1%——
- #109AnthropicClaude Sonnet 410.7%0.3%1404
- #110AnthropicClaude Haiku 4.510.4%0.0%1402
- #111UpstageSolar Pro 310.3%0.0%—
- #112Mistral AIMistral Small 49.9%——
- #113OpenAIGPT-5 Nano9.5%—1348
- #114GoogleGemini 2.5 Flash Lite7.0%—1361
- #115MeituanLongcat Flash Chat5.8%—1434
- #116SpaceXSIGrok 4.1 Fast5.1%——
- #117MetaLlama 4 Maverick4.9%0.0%1317
- #117PerplexitySonar4.9%——
- #119NVIDIANemotron 3 Nano Omni4.8%——
- #120OpenAIGPT-4.14.2%—1375
- #121IBMGranite 4.1 8B3.8%—1318
- #121MetaLlama 4 Scout3.8%0.0%1310
Half of models ≤ 33.6%
49 of the 64 AIs released in the last 60 days have no score here yet · Gemini 4 Neon, GPT-6.2 Astra, Microsoft-Decision-1.1 and 46 more
This score comes from a fixed, predefined evaluation, so it may differ from what you get with your own question.
Scores labelled 'Self-reported' were published by the AI maker itself, not measured by a third party.