Humanity's Last Exam (HLE)
Claude Fable 5.2 leads the HLE leaderboard on BenchLM's November 2026 update with 74.2%, ahead of Kimi K3.1 Pro (70%) and Claude Fable 5.1 (65%), across 81 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 — November 12, 2026
As of November 12, 2026, Claude Fable 5.2 leads the HLE leaderboard with 74.2% , followed by Kimi K3.1 Pro (70%) and Claude Fable 5.1 (65%).
Claude Fable 5.2
Anthropic
Reported by AnthropicKimi K3.1 Pro
Moonshot AI
Reported by Moonshot AIClaude Fable 5.1
Anthropic
Reported by Anthropic81 modelsKnowledgeHLE follows the active BenchAlign v5.8 policy.MethodologyCurrentUpdated November 12, 2026
Leaderboard (81 models)
None of the 81 results on this page was run by the benchmark or an independent evaluator: 77 reported by the labs that built the models and 4 reported by a third party.
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2AnthropicReported by Anthropic | 74.2% | Reported by Anthropic | Not reported | Closed |
| 2 | Kimi K3.1 ProMoonshot AIReported by Moonshot AI | 70% | Reported by Moonshot AI | Not reported | Open |
| 3 | Claude Fable 5.1AnthropicReported by Anthropic | 65% | Reported by Anthropic | Not reported | Closed |
| 4 | Claude Opus 5AnthropicReported by Anthropic | 64.7% | Reported by Anthropic | Not reported | Closed |
| 5 | Claude Mythos 5AnthropicReported by Anthropic | 64.5% | Reported by Anthropic | Not reported | Closed |
| 6 | Muse Spark 1.1MetaReported by Meta | 62.1% | Reported by Meta | Not reported | Closed |
| 7 | Hy4TencentReported by Tencent | 59.1% | Reported by Tencent | Not reported | Pending |
| 8 | GPT-5.4 ProOpenAIReported by OpenAI | 58.7% | Reported by OpenAI | Not reported | Closed |
| 9 | Claude Opus 4.8AnthropicReported by Anthropic | 57.9% | Reported by Anthropic | Not reported | Closed |
| 10 | Claude Sonnet 5AnthropicReported by Anthropic | 57.4% | Reported by Anthropic | Not reported | Closed |
| 11 | GPT-5.5 ProOpenAIReported by OpenAI | 57.2% | Reported by OpenAI | Not reported | Closed |
| 12 | Apodex 1.1ApodexReported by Apodex | 56.1% | Reported by Apodex | Not reported | Closed |
| 13 | Kimi K3Moonshot AIReported by Moonshot AI | 56% | Reported by Moonshot AI | Not reported | Open |
| 14 | Hy4 previewTencentReported by Tencent | 55.4% | Reported by Tencent | Not reported | Open |
| 15 | Claude Opus 4.7 (Adaptive)AnthropicReported by Anthropic | 54.7% | Reported by Anthropic | Not reported | Closed |
| 16 | Inkling 2 SmallThinking Machines LabReported by Thinking Machines Lab | 53.2% | Reported by Thinking Machines Lab | Not reported | Closed |
| 17 | Claude Opus 4.6AnthropicReported by Anthropic | 53% | Reported by Anthropic | Not reported | Closed |
| 18 | dots3-note PreviewDots StudioReported by Dots Studio | 52.6% | Reported by Dots Studio | Not reported | Open |
| 19 | GPT-5.5OpenAIReported by OpenAI | 52.2% | Reported by OpenAI | Not reported | Closed |
| 20 | GPT-5.4OpenAIReported by OpenAI | 52.1% | Reported by OpenAI | Not reported | Closed |
| 21 | Muse SparkMetaReported by Meta | 50.4% | Reported by Meta | Not reported | Closed |
| 22 | Claude Sonnet 4.6AnthropicReported by Anthropic | 49% | Reported by Anthropic | Not reported | Closed |
| 23 | MiMo-V2.5-ProXiaomiReported by Xiaomi | 48% | Reported by Xiaomi | Not reported | Closed |
| 24 | Inkling-SmallThinking Machines LabReported by Thinking Machines Lab | 47.8% | Reported by Thinking Machines Lab | Not reported | Open |
| 25 | Agents-A1InternScienceReported by InternScience | 47.6% | Reported by InternScience | Not reported | Open |
| 26 | Step 5 PreviewStepFunReported by StepFun | 46.5% | Reported by StepFun | Not reported | Open |
| 27 | InklingThinking Machines LabReported by Thinking Machines Lab | 46% | Reported by Thinking Machines Lab | Not reported | Open |
| 28 | Ornith-1.5-397BOrnith AIReported by Ornith AI | 44.6% | Reported by Ornith AI | Not reported | Open |
| 29 | Qwen3.8 MaxAlibabaReported by Alibaba | 43.6% | Reported by Alibaba | Not reported | Open |
| 30 | DeepSeek V4 Pro 0813DeepSeekReported by DeepSeek | 42.7% | Reported by DeepSeek | Not reported | Open |
| 31 | DeepSeek V4.1 Flash 1017DeepSeekReported by DeepSeek | 42.6% | Reported by DeepSeek | Not reported | Open |
| 32 | GPT-5.4 miniOpenAIReported by OpenAI | 41.5% | Reported by OpenAI | Not reported | Closed |
| 33 | Qwen3.7 MaxAlibabaReported by Alibaba | 41.4% | Reported by Alibaba | Not reported | Closed |
| 34 | GLM-5.2Z.AIReported by Z.AI | 40.5% | Reported by Z.AI | Not reported | Open |
| 35 | Gemini 3.5 FlashGoogleReported by Google | 40.2% | Reported by Google | Not reported | Closed |
| 36 | GPT-5.4 nanoOpenAIReported by OpenAI | 37.7% | Reported by OpenAI | Not reported | Closed |
| 37 | Ling 3.1 FlashInclusionAIReported by InclusionAI | 37.4% | Reported by InclusionAI | Not reported | Open |
| 38 | Qwen3.8-Omni-FlashAlibabaReported by Alibaba | 36.5% | Reported by Alibaba | Not reported | Closed |
| 39 | Qwen3.8-Flash-NextAlibabaReported by Alibaba | 35.9% | Reported by Alibaba | Not reported | Open |
| 40 | Grok 4.3xAIReported by a third party | 35% | Reported by a third party | Not reported | Closed |
| 41 | DeepSeek V4 Flash 0731DeepSeekReported by DeepSeek | 34.8% | Reported by DeepSeek | Not reported | Open |
| 42 | Qwen3.7 PlusAlibabaReported by Alibaba | 34.7% | Reported by Alibaba | Not reported | Closed |
| 43 | Kimi K2.6Moonshot AIReported by Moonshot AI | 34.7% | Reported by Moonshot AI | Not reported | Open |
| 44 | DeepSeek V4 Pro (High)DeepSeekReported by DeepSeek | 34.5% | Reported by DeepSeek | Not reported | Open |
| 45 | GLM-5.1Z.AIReported by Z.AI | 31% | Reported by Z.AI | Not reported | Open |
| 46 | Claude Opus 4.5AnthropicAnother lab’s comparison table | 30.8% | Another lab’s comparison table | Not reported | Closed |
| 47 | Qwen3.8-27BAlibabaReported by Alibaba | 30.8% | Reported by Alibaba | Not reported | Open |
| 48 | GLM-5Z.AIReported by Z.AI | 30.5% | Reported by Z.AI | Not reported | Open |
| 49 | Kimi K2.5Moonshot AIReported by Moonshot AI | 30.1% | Reported by Moonshot AI | Not reported | Open |
| 50 | DeepSeek V4 Flash (High)DeepSeekReported by DeepSeek | 29.4% | Reported by DeepSeek | Not reported | Open |
| 51 | Nemotron 3.5 Super VL 120B A12BNVIDIAReported by NVIDIA | 29.2% | Reported by NVIDIA | Not reported | Open |
| 52 | Nemotron 3.5 Super VL 120B A12B NVFP4NVIDIAReported by NVIDIA | 29% | Reported by NVIDIA | Not reported | Open |
| 53 | Qwen3.6 PlusAlibabaReported by Alibaba | 28.8% | Reported by Alibaba | Not reported | Closed |
| 54 | Qwen3.5 397BAlibabaReported by Alibaba | 28.7% | Reported by Alibaba | Not reported | Open |
| 55 | MiniMax M2.7MiniMaxAnother lab’s comparison table | 28% | Another lab’s comparison table | Not reported | Open |
| 56 | A.X K2SK TelecomReported by SK Telecom | 27.8% | Reported by SK Telecom | Not reported | Open |
| 57 | Nemotron 3 UltraNVIDIAReported by NVIDIA | 26.7% | Reported by NVIDIA | Not reported | Open |
| 58 | Gemma 4 31BGoogleReported by Google | 26.5% | Reported by Google | Not reported | Open |
| 59 | Ornith-1.5-35B-A3BOrnith AIReported by Ornith AI | 25.6% | Reported by Ornith AI | Not reported | Open |
| 60 | Hy3 PreviewTencentReported by a third party | 25.5% | Reported by a third party | Not reported | Open |
| 61 | GLM-4.7Z.AIReported by Z.AI | 24.8% | Reported by Z.AI | Not reported | Open |
| 62 | Qwen3.6-27BAlibabaReported by Alibaba | 24% | Reported by Alibaba | Not reported | Open |
| 63 | Ling 3.0 FlashInclusionAIReported by InclusionAI | 22.7% | Reported by InclusionAI | Not reported | Open |
| 64 | Qwen3.6-35B-A3BAlibabaReported by Alibaba | 21.4% | Reported by Alibaba | Not reported | Open |
| 65 | Ornith-1.5-9BOrnith AIReported by Ornith AI | 20.2% | Reported by Ornith AI | Not reported | Open |
| 66 | Gemini 2.5 ProGoogleReported by Google | 18.8% | Reported by Google | Not reported | Closed |
| 67 | K-EXAONE 2.0LG AI ResearchReported by LG AI Research | 18.3% | Reported by LG AI Research | Not reported | Open |
| 68 | Nemotron 3 Super 100BNVIDIAReported by NVIDIA | 18.3% | Reported by NVIDIA | Not reported | Open |
| 69 | Qwen3 235B 2507 (Reasoning)AlibabaReported by Alibaba | 18.2% | Reported by Alibaba | Not reported | Open |
| 70 | DeepSeek-R1-0528DeepSeekReported by DeepSeek | 17.7% | Reported by DeepSeek | Not reported | Open |
| 71 | Gemma 4 26B A4BGoogleReported by Google | 17.2% | Reported by Google | Not reported | Open |
| 72 | DeepSeek V3.1 (Reasoning)DeepSeekReported by DeepSeek | 15.9% | Reported by DeepSeek | Not reported | Open |
| 73 | GPT-OSS 120BOpenAIReported by OpenAI | 14.9% | Reported by OpenAI | Not reported | Open |
| 74 | GPT-OSS 20BOpenAIReported by OpenAI | 10.9% | Reported by OpenAI | Not reported | Open |
| 75 | Nemotron 3 Nano 30BNVIDIAReported by NVIDIA | 10.6% | Reported by NVIDIA | Not reported | Open |
| 76 | MiniCPM5-2BOpenBMBReported by OpenBMB | 8.9% | Reported by OpenBMB | Not reported | Open |
| 77 | DeepSeek V4 FlashDeepSeekReported by DeepSeek | 8.1% | Reported by DeepSeek | Not reported | Open |
| 78 | DeepSeek V4 ProDeepSeekReported by DeepSeek | 7.7% | Reported by DeepSeek | Not reported | Open |
| 79 | LFM2.5-24B-A2BLiquidAIReported by LiquidAI | 7.4% | Reported by LiquidAI | Not reported | Open |
| 80 | LFM3-1.7BLiquidAIReported by LiquidAI | 7.2% | Reported by LiquidAI | Not reported | Open |
| 81 | Kimi K2Moonshot AIReported by Moonshot AI | 4.7% | Reported by Moonshot AI | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
Results reported by model developersUnranked
Labs ran these themselves, with a metric, split or setting the table above does not use. They take no rank here.
| Model | Reported value | What was measured | Against the table above | Who ran it |
|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017DeepSeek | 44.0%Accuracy | HLE w/o tools, text-only subsetText-only subset; the stored score is the full set | Not comparable to the scores above | Reported by DeepSeek |
A lab’s own result is shown for reference. It never takes a rank on this board.
How we label sourcesHow 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: 81 sourced rows are currently displayable on this page; the leading published row is Claude Fable 5.2 at 74.2%.
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, Claude Fable 5.2 leads the HLE benchmark with a score of 74.2%, followed by Kimi K3.1 Pro (70%) and Claude Fable 5.1 (65%). The scores show moderate spread, with meaningful differences between the top tier and mid-tier models.
81 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 7,000+ subscribers.
One email each week. Unsubscribe anytime.