Benchmark leaderboard - 25 entries
Humanity's Last Exam leaderboard
Humanity's Last Exam is a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. As of October 17, 2026, GPT 6 Astra ranks first on the Humanity's Last Exam board with 54.80±1.94.
Last verified from labs.scale.com. Read the caveats
Humanity's Last Exam
25 entries, in the board's order
| 1 | GPT 6 Astra Scale Labs SEAL leaderboard listed position 1; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 39; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 54.80±1.94 |
| 2 | Fable 5.1 (xhigh) Scale Labs SEAL leaderboard listed position 2; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 20; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 46.50±2.00 |
| 3 | gemini-3.1-pro-preview (thinking high) Scale Labs SEAL leaderboard listed position 3; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 51; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 46.44±1.96 |
| 4 | Sonnet 5.5 (xhigh) Scale Labs SEAL leaderboard listed position 4; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error not shown; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 46.12±1.95 |
| 5 | Gemini 3.8 Flash Scale Labs SEAL leaderboard listed position 5; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 51; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 44.52±1.96 |
| 6 | gpt-5.4-pro-2026-03-05 Scale Labs SEAL leaderboard listed position 6; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 38; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 44.32±1.95 |
| 7 | Muse Spark Scale Labs SEAL leaderboard listed position 7; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 50; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 40.56±1.92 |
| 8 | gemini-3-pro-preview Scale Labs SEAL leaderboard listed position 8; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 57; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 37.52±1.90 |
| 9 | gpt-5.4-2026-03-05 (xhigh thinking) Scale Labs SEAL leaderboard listed position 9; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 42; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 36.24±1.88 |
| 10 | claude-opus-4-7 Scale Labs SEAL leaderboard listed position 10; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 47; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 36.20±1.88 |
| 11 | claude-opus-4-6-thinking-max Scale Labs SEAL leaderboard listed position 11; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 46; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 34.44±1.86 |
| 12 | gpt-5-pro-2025-10-06 Scale Labs SEAL leaderboard listed position 12; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 49; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 31.64±1.82 |
| 13 | Claude Haiku 5.5 Scale Labs SEAL leaderboard listed position 13; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error not shown; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 30.12±1.83 |
| 14 | gpt-5.2-2025-12-11 Scale Labs SEAL leaderboard listed position 14; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 45; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 27.80±1.76 |
| 15 | gpt-5-2025-08-07 Scale Labs SEAL leaderboard listed position 15; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 50; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 25.32±1.70 |
| 16 | claude-opus-4-5-20251101-thinking Scale Labs SEAL leaderboard listed position 16; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 55; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 25.20±1.70 |
| 17 | kimi-k2.5 Scale Labs SEAL leaderboard listed position 17; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 67; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 24.37±1.81 |
| 18 | gpt-5.1-thinking Scale Labs SEAL leaderboard listed position 18; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 55; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 23.68±1.67 |
| 19 | gemini-2.5-pro-preview-06-05 Scale Labs SEAL leaderboard listed position 19; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 72; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 21.64±1.61 |
| 20 | o3 (high) (April 2025) Scale Labs SEAL leaderboard listed position 20; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 34; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 20.32±1.58 |
| 21 | gpt-5-mini-2025-08-07 Scale Labs SEAL leaderboard listed position 21; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 65; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 19.44±1.55 |
| 22 | o3 (medium) (April 2025) Scale Labs SEAL leaderboard listed position 22; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 39; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 19.20±1.54 |
| 23 | claude-opus-4-6 (Non-Thinking) Scale Labs SEAL leaderboard listed position 23; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 44; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 19.00±1.54 |
| 24 | Gemini 2.5 Pro Experimental (March 2025) Scale Labs SEAL leaderboard listed position 24; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 71; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 18.16±1.51 |
| 25 | o4-mini (high) (April 2025) Scale Labs SEAL leaderboard listed position 25; HLE public set (2,500 questions), temperature 0.0 when configurable, o3-mini-2025-01-31 judge; score is accuracy % with 95% CI; calibration error 57; no per-entry eval date shown, date is leaderboard read date Official leaderboard | 18.08±1.51 |
About this board
- Per-entry evaluation dates: the official Scale Labs HLE leaderboard (labs.scale.com/leaderboard/humanitys_last_exam, linked from lastexam.ai) shows no per-entry dates, so every entry uses the read date 2026-10-17.
- Calibration error for Sonnet 5.5 (xhigh) and Claude Haiku 5.5: not shown on the leaderboard.
- HLE-Rolling and HLE-Diamond scores: lastexam.ai charts them via the CAIS dashboard (dashboard.safe.ai) but publishes no table on the page, so they are not recorded.
- lastexam.ai's own results table (judge o3-mini, dataset dated April 3rd, 2025) lists only 10 older models, topped by Gemini 3 Pro at 38.3; it was not used because the Scale Labs leaderboard is more current.
FAQ
What does Humanity's Last Exam measure?
Humanity's Last Exam is a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage.
Which model leads Humanity's Last Exam?
As of October 17, 2026, GPT 6 Astra ranks first on the Humanity's Last Exam board with 54.80±1.94. That is the board's own order, not a Developers Digest ranking.
Where do these Humanity's Last Exam scores come from?
All 25 entries come from the official leaderboard at labs.scale.com. Every entry links to its source, which was read on October 17, 2026, and lists the setting it was run under.
Related reading
- Claude Haiku 5.5: Pricing, Migration Changes, and When to Use ItClaude Haiku 5.5 pricing, the cliff above 100K-token prompts, the tokenizer catch, the migration steps from Haiku 4.5, and when to use it instead of Sonnet.
- Claude Sonnet 5.5 Developer Guide: Pricing and API ChangesClaude Sonnet 5.5 costs $2/$10 per million tokens, and cache reads fell to $0.10 on October 7 - half of Opus 5.5 on every line. Pricing math and API changes.
- Claude Opus 5.5 Developer Guide: Pricing, API, Claude CodeClaude Opus 5.5 costs $4/$20 per million tokens with $0.20 cache reads. SDK examples, Claude Code setup, and when to pick it over Sonnet 5.5 or Haiku 5.5.
- Claude Mythos Preview Explained: Anthropic's Gated Frontier Model and Project GlasswingClaude Mythos Preview explained: Anthropic's gated frontier model, who got access through Project Glasswing, what it found, and where the model line went next.
- Claude Opus 5 vs Opus 4.8 vs Fable 5: Benchmark Comparison (July 2026)Claude Opus 5 launched July 24, 2026 at $5/$25 per MTok - matching Opus 4.8 pricing while delivering near-Fable 5 intelligence. Full benchmark comparison across 7 evals, pricing breakdown, and decision guide.
- Inkling: Thinking Machines Lab Drops a 975B Open-Weights ModelA new American open-weights frontier model with multimodal capabilities, 1M token context, and competitive benchmarks. Here's what the HN community thinks.
Sources
Last verified . Each row links to the page its score was copied from.
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.