// BENCHMARK

Humanity's Last Exam benchmark

AI model leaderboard for the Humanity's Last Exam benchmark. Compare how large language models score on Humanity's Last Exam, see the full ranking, and understand what this AI benchmark measures. Kimi K2 Thinking currently leads with 51%. ~3,000 expert-curated questions across 100+ subjects designed to remain hard for frontier AI.

Leaderboard

# Model Organization Score Variant Source
#1 Kimi K2 Thinking Moonshot 51% thinking-heavy official ↗
#2 Kimi K2.6 Moonshot 50.2% with-tools official ↗
#3 Kimi K2.5 Moonshot 50.2% with-tools official ↗
#4 Sakana Fugu Sakana AI 47.2% official ↗
#5 Gemini 3 Pro Google DeepMind 45.8% search-agent-cited-deepseek-v3-2 official ↗
#6 Kimi K2 Thinking Moonshot 44.9% thinking-search-agent-cited-deepseek-v3-2 official ↗
#7 Kimi K2 Thinking Moonshot 44.9% thinking-w-tools official ↗
#8 Claude Opus 4.5 Anthropic 43.2% with-search official ↗
#9 DeepSeek-V3.2 DeepSeek 40.8% thinking-search-agent official ↗
#10 GLM-5.2 Zhipu AI 40.5%
#11 Gemini 3.5 Flash Google DeepMind 40.2% official ↗
#12 Tongyi DeepResearch Alibaba-NLP 38.3% heavy-mode official ↗
#13 Qwen3.7-Max Qwen 38.1% official ↗
#14 Gemini 3 Pro Google DeepMind 37.7% text-only-cited-deepseek-v3-2 official ↗
#15 GPT-5 OpenAI 35.2% search-agent-cited-deepseek-v3-2 official ↗
#16 Grok 4.3 xAI 35% official ↗
#17 GPT-5.5 Pro OpenAI 34.1% parallel-compute official ↗
#18 Tongyi DeepResearch Alibaba-NLP 32.9% react-avg@3 official ↗
#19 Claude Sonnet 4.5 Anthropic 32% search-agent-cited-deepseek-v3-2 official ↗
#20 MiniMax-M2 MiniMax 31.8% search-agent-cited-deepseek-v3-2 official ↗
#21 DeepSeek-V3.1 DeepSeek 29.8% react-cited-tongyi official ↗
#22 GPT-5 OpenAI 28.63% High official ↗
#23 GPT-5.5 OpenAI 28.4% no-tools official ↗
#24 NVIDIA Nemotron 3 Ultra NVIDIA 26.7% official ↗
#25 GPT-5 OpenAI 26.3% text-only-cited-deepseek-v3-2 official ↗
#26 DeepSeek-V3.2 DeepSeek 25.1% thinking-text-only official ↗
#27 o3 OpenAI 24.9% react-cited-tongyi official ↗
#28 GPT-5 OpenAI 24.84% thinking-high-cited-ring-1t official ↗
#29 GPT-5.4 Thinking OpenAI 24.8% thinking official ↗
#30 Kimi K2 Thinking Moonshot 23.9% thinking-no-tools official ↗
#31 Grok 4 xAI 23.9% official ↗
#32 Kimi K2 Thinking Moonshot 23.9% thinking-text-only-cited-deepseek-v3-2 official ↗
#33 o3 OpenAI 22.34% High official ↗
#34 Kimi K2 Moonshot 21.7% k2-0905-w-tools-cited-k2-thinking official ↗
#35 gemini-2.5-pro Google DeepMind 21.6% cited-ring-1t official ↗
#36 GLM 4.5 Zhipu AI 21.2% react-cited-tongyi official ↗
#37 gemini-2.5-pro Google DeepMind 21.1% official ↗
#38 Claude Sonnet 4 Anthropic 20.3% react-cited-tongyi official ↗
#39 o3 OpenAI 20% official ↗
#40 gemini-2.5-pro Google DeepMind 19.93% official ↗
#41 DeepSeek-V3.2-Exp DeepSeek 19.8% from-hf-readme official ↗
#42 gpt-oss-120b OpenAI 19% high-with-tools official ↗
#43 Kimi K2 Moonshot 18.1% react-cited-tongyi official ↗
#44 DeepSeek-V3.1-Terminus DeepSeek 17.82% thinking-cited-ring-1t official ↗
#45 o4-mini OpenAI 17.7% react-cited-tongyi official ↗
#46 gpt-oss-20b OpenAI 17.3% high-with-tools official ↗
#47 Ring-1T Ant Group 16.03% thinking official ↗
#48 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 15.8% official ↗
#49 gpt-oss-120b OpenAI 14.9% high reasoning, no tools official ↗
#50 DeepSeek-R1-0528 DeepSeek 14.9% official ↗
#51 gpt-oss-120b OpenAI 14.9% high-no-tools official ↗
#52 GLM 4.5 Zhipu AI 14.4% official ↗
#53 Qwen3-235B-A22B Qwen 14.23% official ↗
#54 Qwen3-235B-A22B-Thinking (Jul 2025) Qwen 13.95% thinking-cited-ring-1t official ↗
#55 Claude Sonnet 4.5 Anthropic 13.7% text-only-cited-deepseek-v3-2 official ↗
#56 MiniMax-M2 MiniMax 12.5% text-only-cited-deepseek-v3-2 official ↗
#57 OpenReasoning-Nemotron-32B NVIDIA 12.26% official ↗
#58 gemini-2.5-pro Google DeepMind 12.07% lowthink-cited-ling-1t official ↗
#59 Claude Opus 4 Anthropic 11.7% official ↗
#60 gpt-oss-20b OpenAI 11.23% official ↗
#61 gpt-oss-20b OpenAI 10.9% high-no-tools official ↗
#62 gpt-oss-20b OpenAI 10.9% high reasoning, no tools official ↗
#63 GLM-4.5-Air Zhipu AI 10.6% official ↗
#64 DeepSeek-V3.1-Terminus DeepSeek 10.38% non-thinking-cited-ling-1t official ↗
#65 QwQ-32B Qwen 9.98% official ↗
#66 K2 Think Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) 9.95% avg@16 official ↗
#67 Magistral Medium 1.1 Mistral AI 9% text-only official ↗
#68 DeepSeek-R1 DeepSeek 8.6% text-only-cited-magistral official ↗
#69 DeepSeek-R1 DeepSeek 8.5% official ↗
#70 DeepSeek-V3.1 DeepSeek 8.4% official ↗
#71 MiniMax-M1-80k MiniMax 8.4% no-tools official ↗
#72 Kimi K2 Moonshot 7.9% k2-0905-no-tools-cited-k2-thinking official ↗
#73 Llama Nemotron Super 49B NVIDIA 7.64% reasoning-on-text-only-pass@1 official ↗
#74 Ling-1T Ant Group 7.6% pass@1 official ↗
#75 GPT-5 OpenAI 7.33% main-cited-ling-1t official ↗
#76 Kimi K2 Moonshot 7.29% instruct-0905-cited-ling-1t official ↗
#77 MiniMax-M1-40k MiniMax 7.2% no-tools official ↗
#78 NVIDIA-Nemotron-Nano-9B-v2 NVIDIA 6.5% from-hf-readme official ↗
#79 Magistral Small 1.1 Mistral AI 6.4% text-only official ↗
#80 Qwen3-30B-A3B Qwen 6.14% official ↗
#81 Mistral Medium 3 Mistral AI 4.4% text-only-cited-magistral official ↗

Frequently asked questions about Humanity's Last Exam

What is the Humanity's Last Exam benchmark?

~3,000 expert-curated questions across 100+ subjects designed to remain hard for frontier AI.

How is the Humanity's Last Exam benchmark scored?

Humanity's Last Exam is scored using the accuracy (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.

Which AI model scores highest on Humanity's Last Exam?

As of the latest reported scores on GenAIList, Kimi K2 Thinking achieves the highest result on Humanity's Last Exam with a score of 51%.

Is a higher Humanity's Last Exam score better?

Yes. On Humanity's Last Exam a higher score indicates better performance, so models near the top of the leaderboard are the strongest.