Humanity's Last Exam (HLE)
Kimi K3.1 Pro leads the HLE leaderboard on BenchLM's October 2026 update with 70%, ahead of Claude Fable 5.1 (65%) and Claude Opus 5 (64.7%), across 77 models.
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.
Top models on HLE — October 23, 2026
As of October 23, 2026, Kimi K3.1 Pro leads the HLE leaderboard with 70% , followed by Claude Fable 5.1 (65%) and Claude Opus 5 (64.7%).
Kimi K3.1 Pro
Moonshot AI
Claude Fable 5.1
Anthropic
Claude Opus 5
Anthropic
77 modelsKnowledgeHLE follows the active BenchAlign v5.8 policy.MethodologyCurrentUpdated October 23, 2026
Leaderboard (77 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Kimi K3.1 ProMoonshot AI | 70% | Not reported | Pending |
| 2 | Claude Fable 5.1Anthropic | 65% | Not reported | Closed |
| 3 | Claude Opus 5Anthropic | 64.7% | Not reported | Closed |
| 4 | Claude Mythos 5Anthropic | 64.5% | Not reported | Closed |
| 5 | Muse Spark 1.1Meta | 62.1% | Not reported | Closed |
| 6 | GPT-5.4 ProOpenAI | 58.7% | Not reported | Closed |
| 7 | Claude Opus 4.8Anthropic | 57.9% | Not reported | Closed |
| 8 | Claude Sonnet 5Anthropic | 57.4% | Not reported | Closed |
| 9 | GPT-5.5 ProOpenAI | 57.2% | Not reported | Closed |
| 10 | Apodex 1.1Apodex | 56.1% | Not reported | Closed |
| 11 | Kimi K3Moonshot AI | 56% | Not reported | Pending |
| 12 | Hy4 previewTencent | 55.4% | Not reported | Open |
| 13 | Claude Opus 4.7 (Adaptive)Anthropic | 54.7% | Not reported | Closed |
| 14 | Claude Opus 4.6Anthropic | 53% | Not reported | Closed |
| 15 | dots3-note PreviewDots Studio | 52.6% | Not reported | Open |
| 16 | GPT-5.5OpenAI | 52.2% | Not reported | Closed |
| 17 | GPT-5.4OpenAI | 52.1% | Not reported | Closed |
| 18 | Muse SparkMeta | 50.4% | Not reported | Closed |
| 19 | Claude Sonnet 4.6Anthropic | 49% | Not reported | Closed |
| 20 | MiMo-V2.5-ProXiaomi | 48% | Not reported | Closed |
| 21 | Inkling-SmallThinking Machines Lab | 47.8% | Not reported | Open |
| 22 | Agents-A1InternScience | 47.6% | Not reported | Open |
| 23 | Step 5 PreviewStepFun | 46.5% | Not reported | Open |
| 24 | InklingThinking Machines Lab | 46% | Not reported | Open |
| 25 | Ornith-1.5-397BOrnith AI | 44.6% | Not reported | Open |
| 26 | Qwen3.8 MaxAlibaba | 43.6% | Not reported | Open |
| 27 | DeepSeek V4 Pro 0813DeepSeek | 42.7% | Not reported | Open |
| 28 | DeepSeek V4.1 Flash 1017DeepSeek | 42.6% | Not reported | Open |
| 29 | GPT-5.4 miniOpenAI | 41.5% | Not reported | Closed |
| 30 | Qwen3.7 MaxAlibaba | 41.4% | Not reported | Closed |
| 31 | GLM-5.2Z.AI | 40.5% | Not reported | Open |
| 32 | Gemini 3.5 FlashGoogle | 40.2% | Not reported | Closed |
| 33 | GPT-5.4 nanoOpenAI | 37.7% | Not reported | Closed |
| 34 | Ling 3.1 FlashInclusionAI | 37.4% | Not reported | Open |
| 35 | Qwen3.8-Omni-FlashAlibaba | 36.5% | Not reported | Closed |
| 36 | Qwen3.8-Flash-NextAlibaba | 35.9% | Not reported | Open |
| 37 | Grok 4.3xAI | 35% | Not reported | Closed |
| 38 | DeepSeek V4 Flash 0731DeepSeek | 34.8% | Not reported | Open |
| 39 | Qwen3.7 PlusAlibaba | 34.7% | Not reported | Closed |
| 40 | Kimi K2.6Moonshot AI | 34.7% | Not reported | Open |
| 41 | DeepSeek V4 Pro (High)DeepSeek | 34.5% | Not reported | Open |
| 42 | GLM-5.1Z.AI | 31% | Not reported | Open |
| 43 | Claude Opus 4.5Anthropic | 30.8% | Not reported | Closed |
| 44 | Qwen3.8-27BAlibaba | 30.8% | Not reported | Open |
| 45 | GLM-5Z.AI | 30.5% | Not reported | Open |
| 46 | Kimi K2.5Moonshot AI | 30.1% | Not reported | Open |
| 47 | DeepSeek V4 Flash (High)DeepSeek | 29.4% | Not reported | Open |
| 48 | Nemotron 3.5 Super VL 120B A12BNVIDIA | 29.2% | Not reported | Open |
| 49 | Nemotron 3.5 Super VL 120B A12B NVFP4NVIDIA | 29% | Not reported | Open |
| 50 | Qwen3.6 PlusAlibaba | 28.8% | Not reported | Closed |
| 51 | Qwen3.5 397BAlibaba | 28.7% | Not reported | Open |
| 52 | MiniMax M2.7MiniMax | 28% | Not reported | Open |
| 53 | A.X K2SK Telecom | 27.8% | Not reported | Open |
| 54 | Nemotron 3 UltraNVIDIA | 26.7% | Not reported | Open |
| 55 | Gemma 4 31BGoogle | 26.5% | Not reported | Open |
| 56 | Ornith-1.5-35B-A3BOrnith AI | 25.6% | Not reported | Open |
| 57 | Hy3 PreviewTencent | 25.5% | Not reported | Open |
| 58 | GLM-4.7Z.AI | 24.8% | Not reported | Open |
| 59 | Qwen3.6-27BAlibaba | 24% | Not reported | Open |
| 60 | Ling 3.0 FlashInclusionAI | 22.7% | Not reported | Open |
| 61 | Qwen3.6-35B-A3BAlibaba | 21.4% | Not reported | Open |
| 62 | Ornith-1.5-9BOrnith AI | 20.2% | Not reported | Open |
| 63 | Gemini 2.5 ProGoogle | 18.8% | Not reported | Closed |
| 64 | K-EXAONE 2.0LG AI Research | 18.3% | Not reported | Open |
| 65 | Nemotron 3 Super 100BNVIDIA | 18.3% | Not reported | Open |
| 66 | Qwen3 235B 2507 (Reasoning)Alibaba | 18.2% | Not reported | Open |
| 67 | DeepSeek-R1-0528DeepSeek | 17.7% | Not reported | Open |
| 68 | Gemma 4 26B A4BGoogle | 17.2% | Not reported | Open |
| 69 | DeepSeek V3.1 (Reasoning)DeepSeek | 15.9% | Not reported | Open |
| 70 | GPT-OSS 120BOpenAI | 14.9% | Not reported | Open |
| 71 | GPT-OSS 20BOpenAI | 10.9% | Not reported | Open |
| 72 | Nemotron 3 Nano 30BNVIDIA | 10.6% | Not reported | Open |
| 73 | MiniCPM5-2BOpenBMB | 8.9% | Not reported | Open |
| 74 | DeepSeek V4 FlashDeepSeek | 8.1% | Not reported | Open |
| 75 | DeepSeek V4 ProDeepSeek | 7.7% | Not reported | Open |
| 76 | LFM2.5-24B-A2BLiquidAI | 7.4% | Not reported | Open |
| 77 | Kimi K2Moonshot AI | 4.7% | Not reported | Closed |
How to read this leaderboard
The finalized stable Humanity's Last Exam contains 2,500 difficult, closed-ended questions across more than 100 subjects. It combines exact-match, multiple-choice and image-dependent items to probe expert-level academic question answering.
Variant check: Stable HLE, HLE-Rolling, HLE-Verified, and HLE-Diamond use different item sets. At the September 22, 2026 editorial check, the source ledger mixed rows reported with tools and without tools. The table does not expose a complete protocol label for every row.
Compare HLE results only after matching tool access and the evaluation setup. Consult the original provider or benchmark report because this table does not expose a complete per-row protocol. Search, browsing or code execution changes the task; a tool-assisted result is not a controlled comparison with a closed-book run.
Operator receipt: 77 sourced rows are currently displayable on this page; the leading published row is Kimi K3.1 Pro at 70%.
Provenance: At the September 22, 2026 editorial check, the official stable-set leaderboard was not the source of this page's full model table. The source ledger contained provider and secondary reports, with no benchmark-owner or independent result rows.
Source freshness: The stable set was finalized on April 3, 2025, and the official leaderboard names that date as its dataset update. HLE-Rolling launched on October 8, 2025. The peer-reviewed Nature article was published on January 28, 2026.
Honest limit: HLE measures difficult closed-ended academic question answering, not autonomous research ability or artificial general intelligence. The final paper also reports stronger mathematics and STEM representation. Treat small gaps across unlike or incompletely disclosed protocols as directional, not decisive.
According to BenchLM.ai, Kimi K3.1 Pro leads the HLE benchmark with a score of 70%, followed by Claude Fable 5.1 (65%) and Claude Opus 5 (64.7%). The scores show moderate spread, with meaningful differences between the top tier and mid-tier models.
77 models have been evaluated on HLE. The benchmark falls in the Knowledge category. HLE follows the active BenchAlign v5.8 policy. See the methodology for the current scoring and qualification rules.
About HLE
Year
2025
Tasks
Expert-level questions
Format
Open-ended and multiple choice
Difficulty
Frontier expert level
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.
Freshness and provenance
Version
Humanity's Last Exam
Refresh cadence
Static
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does Humanity's Last Exam measure?
Humanity's Last Exam measures answers to 2,500 difficult, closed-ended academic questions across more than 100 subjects. The finalized set includes image-dependent, multiple-choice, and exact-match items. It is evidence about expert-level question answering under a stated protocol, not proof of autonomous research ability or artificial general intelligence.
Can I compare HLE scores with and without tools?
Only after matching the evaluation setup. The table does not expose a complete protocol label for every row, so consult the original provider or benchmark report before comparing results. A model with search, browsing, code execution, or another tool has not taken the same test as a closed-book model.
How do stable HLE, HLE-Rolling, HLE-Verified, and HLE-Diamond differ?
The stable HLE set contains 2,500 questions. HLE-Rolling is maintained separately, HLE-Verified checks and revises original items, and HLE-Diamond selects 1,000 refined questions split evenly between reasoning and knowledge. Each has a different item set, so a score on one should not be treated as a score on another.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.