Humanity's Last Exam
Multi-domain expert-level exam covering science, math, humanities, and more
Models ranked
47
tracked on this benchmark
Score band
64.5 – 6.5
best → lowest tracked
Snapshot trend
-1.48
Oct 17 → Oct 26 · 4 models
Leaderboard
Tracked models ranked by Score (higher is better).
Notes: Official Anthropic launch table starred the 64.5% HLE with-tools score as Claude Mythos 5. The 59.0% no-tools variant is recorded but not seeded because modelBenchmark rows are unique by modelSlug and benchmarkSlug.
Notes: Moonshot self-reported launch table, row 'HLE-Full', column 'Kimi K3.1 Pro (max)'. Settings: reasoning effort max, temperature 1.0, top-p 0.95 (blog footnote). Full 2,500-question set, no tools.
Notes: Model variant: Kimi K3 (max reasoning effort). Benchmark variant: HLE-Full without tools. Harness/evaluator not disclosed by Moonshot. Source posture: Moonshot/Kimi self-reported launch table; confidence: medium because the value is vendor-reported and not independently reproduced. Recommended seed value: 43.5.
Configuration: DeepSeek-V4-Pro-0813 (max effort)
Notes: DeepSeek first-party HF card DeepSeek-V4-Pro-0813, benchmark table (column 1 = DeepSeek-V4-Pro-0813). Note 1: code-agent tasks run with DeepSeek Harness minimal mode, max effort, temperature 1.0, top_p 0.95. Row 'HLE (wo / w tools)': 42.7 / 60.0. Filed under hle (not hle-no-tools) to match the DeepSeek-family precedent (deepseek-v4.1-flash, -1017 rows). Retrieved 2026-10-26.
Configuration: DeepSeek-V4-Flash-0731 (max effort)
Notes: DeepSeek first-party HF card DeepSeek-V4-Pro-0813, benchmark table column 2 = DeepSeek-V4-Flash-0731, row 'HLE (wo / w tools)': 37.8 / 51.5. The Flash-0731 card itself has no HLE row. Filed under hle per DeepSeek-family precedent. Retrieved 2026-10-26.
Notes: Full HLE set, no tools. DeepSeek also reports 39.1 on the HLE text-only subset (base HF card † footnote; changelog 2026-09-10). No separate score row for the text-only subset: the benchmark catalog has hle, hle-no-tools, hle-with-tools, and hle-with-tools-pro only. Rows are unique by modelSlug and benchmarkSlug. Relabeled 2026-10-17.
Configuration: tool setting not stated; GPT-4o judge
Notes: Qwen Team tech report 'Qwen3.8-Omni: Towards Native Omni-Modal Agents' (arXiv 2609.25611v1, 2026-09-22), Table 2 column 'Qwen3.8-Omni-Flash', row 'HLE': 36.5. Tool setting not stated, so filed under generic hle. Retrieved 2026-10-26.
Configuration: tool setting not stated; GPT-4o judge
Notes: Qwen Team tech report 'Qwen3.8-Omni: Towards Native Omni-Modal Agents' (arXiv 2609.25611v1, 2026-09-22), Table 2 column 'Qwen3.8-Flash', row 'HLE': 35.9. Tool setting not stated, so filed under generic hle. Identity: the report's 'Qwen3.8-Flash' column and the QwenCloud qwen3.8-flash docs page carry the same values as the Qwen3.8-Flash-Next blog/HF card; Qwen's blog says Qwen3.8-Flash is the production served version (1M context, built-in tools) while these evals used a 256K context. Retrieved 2026-10-26.
How to read this benchmark
This benchmark scores models where higher is better. Use scores for directional filtering and shortlisting, not universal quality ranking; then validate pricing, context window, provider availability, and fit for your workload.
Trust this score when
- There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
- The model list covers the same version family you can actually deploy today.
- Top candidates overlap with your required routing and feature requirements.
Be cautious when
- There is only one benchmark snapshot or the dataset appears stale.
- The benchmark metric direction is opposite of your decision objective.
- The score difference between options is narrow and likely within implementation variance.
Related benchmarks
Last reviewed: May 20, 2026