# Humanity's Last Exam (HLE) > An expert-authored benchmark designed to probe frontier knowledge and reasoning. BenchLM keeps protocol differences visible because tool-assisted and closed-book HLE runs answer different questions. Canonical page: https://benchlm.ai/benchmarks/hle - Category: [Knowledge](/knowledge) - Last updated: September 4, 2026 ## About HLE - Year: 2025 - Tasks: Expert-level questions - Format: Open-ended and multiple choice - Difficulty: Frontier expert level - Paper: [Humanity's Last Exam](https://lastexam.ai/) Questions span advanced mathematics, theoretical physics, philosophy, and other specialist fields. The benchmark still has substantial headroom, but its public results increasingly mix tool-assisted and closed-book-style protocols. HLE is currently weighted in BenchLM's scoring formula. The Knowledge category carries 12% of the overall score, and HLE contributes 35% of that category score. ## Leaderboard (53 models) | Rank | Model | Creator | Score | |------|-------|---------|-------| | 1 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 65% | | 2 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 64.7% | | 3 | [Claude Mythos 5](/models/claude-mythos-5) | Anthropic | 64.5% | | 4 | [Muse Spark 1.1](/models/muse-spark-1-1) | Meta | 62.1% | | 5 | [GPT-5.4 Pro](/models/gpt-5-4-pro) | OpenAI | 58.7% | | 6 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 57.9% | | 7 | [Claude Sonnet 5](/models/claude-sonnet-5) | Anthropic | 57.4% | | 8 | [GPT-5.5 Pro](/models/gpt-5-5-pro) | OpenAI | 57.2% | | 9 | [Apodex 1.1](/models/apodex-1-1) | Apodex | 56.1% | | 10 | [Kimi K3](/models/kimi-k3) | Moonshot AI | 56% | | 11 | [Hy4 preview](/models/hy4-preview) | Tencent | 55.4% | | 12 | [Claude Opus 4.7 (Adaptive)](/models/claude-opus-4-7-adaptive) | Anthropic | 54.7% | | 13 | [GLM-5.2](/models/glm-5-2) | Z.AI | 54.7% | | 14 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 53% | | 15 | [dots3-note Preview](/models/dots3-note-preview) | Dots Studio | 52.6% | | 16 | [GLM-5.1](/models/glm-5-1) | Z.AI | 52.3% | | 17 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 52.2% | | 18 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 52.1% | | 19 | [Muse Spark](/models/muse-spark) | Meta | 50.4% | | 20 | [GLM-5](/models/glm-5) | Z.AI | 50.4% | | 21 | [Claude Sonnet 4.6](/models/claude-sonnet-4-6) | Anthropic | 49% | | 22 | [MiMo-V2.5-Pro](/models/mimo-v2-5-pro) | Xiaomi | 48% | | 23 | [Inkling-Small](/models/inkling-small) | Thinking Machines Lab | 47.8% | | 24 | [Agents-A1](/models/agents-a1) | InternScience | 47.6% | | 25 | [Inkling](/models/inkling) | Thinking Machines Lab | 46% | | 26 | [Ornith-1.5-397B](/models/ornith-1-5-397b) | Ornith AI | 44.6% | | 27 | [Qwen3.8 Max](/models/qwen3-8-max) | Alibaba | 43.6% | | 28 | [DeepSeek V4 Pro 0813](/models/deepseek-v4-pro-0813) | DeepSeek | 42.7% | | 29 | [GPT-5.4 mini](/models/gpt-5-4-mini) | OpenAI | 41.5% | | 30 | [Qwen3.7 Max](/models/qwen3-7-max) | Alibaba | 41.4% | | 31 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Google | 40.2% | | 32 | [GPT-5.4 nano](/models/gpt-5-4-nano) | OpenAI | 37.7% | | 33 | [Qwen3.8-Flash-Next](/models/qwen3-8-flash-next) | Alibaba | 35.9% | | 34 | [Grok 4.3](/models/grok-4-3) | xAI | 35% | | 35 | [DeepSeek V4 Flash 0731](/models/deepseek-v4-flash-0731) | DeepSeek | 34.8% | | 36 | [Kimi K2.6](/models/kimi-2-6) | Moonshot AI | 34.7% | | 37 | [Qwen3.7 Plus](/models/qwen3-7-plus) | Alibaba | 34.7% | | 38 | [Qwen3.8-27B](/models/qwen3-8-27b) | Alibaba | 30.8% | | 39 | [Claude Opus 4.5](/models/claude-opus-4-5) | Anthropic | 30.8% | | 40 | [Kimi K2.5](/models/kimi-k2-5) | Moonshot AI | 30.1% | | 41 | [Qwen3.6 Plus](/models/qwen3-6-plus) | Alibaba | 28.8% | | 42 | [Qwen3.5 397B](/models/qwen3-5-397b) | Alibaba | 28.7% | | 43 | [Nemotron 3 Ultra](/models/nemotron-3-ultra) | NVIDIA | 26.7% | | 44 | [Gemma 4 31B](/models/gemma-4-31b) | Google | 26.5% | | 45 | [Ornith-1.5-35B-A3B](/models/ornith-1-5-35b-a3b) | Ornith AI | 25.6% | | 46 | [Hy3 Preview](/models/hy3-preview) | Tencent | 25.5% | | 47 | [GLM-4.7](/models/glm-4-7) | Z.AI | 24.8% | | 48 | [Qwen3.6-27B](/models/qwen3-6-27b) | Alibaba | 24% | | 49 | [Ling 3.0 Flash](/models/ling-3-0-flash) | InclusionAI | 22.7% | | 50 | [Qwen3.6-35B-A3B](/models/qwen3-6-35b-a3b) | Alibaba | 21.4% | | 51 | [Ornith-1.5-9B](/models/ornith-1-5-9b) | Ornith AI | 20.2% | | 52 | [Gemini 2.5 Pro](/models/gemini-2-5-pro) | Google | 18.8% | | 53 | [Gemma 4 26B A4B](/models/gemma-4-26b-a4b) | Google | 17.2% | ## FAQ ### What does HLE measure? An expert-authored benchmark designed to probe frontier knowledge and reasoning. BenchLM keeps protocol differences visible because tool-assisted and closed-book HLE runs answer different questions. ### Which model scores highest on HLE? Claude Fable 5.1 by Anthropic currently leads with a score of 65% on HLE. ### How many models are evaluated on HLE? 53 AI models have been evaluated on HLE on BenchLM. ### Does HLE affect BenchLM's overall score? Yes. HLE is a weighted benchmark inside the Knowledge category, which carries 12% of BenchLM's overall score. HLE itself contributes 35% of that category score. ## Compare Top Models on HLE - [Claude Fable 5.1 vs Claude Opus 5](/compare/claude-fable-5-1-vs-claude-opus-5) - [Claude Opus 5 vs Claude Mythos 5](/compare/claude-mythos-5-vs-claude-opus-5) - [Claude Mythos 5 vs Muse Spark 1.1](/compare/claude-mythos-5-vs-muse-spark-1-1) - [Muse Spark 1.1 vs GPT-5.4 Pro](/compare/gpt-5-4-pro-vs-muse-spark-1-1) ## Related Reading - [HLE benchmark explainer](/blog/posts/hle-humanitys-last-exam)