Share promptsSign inExplore benchmarks
← Reasoning & knowledge benchmarks

Humanity's Last Exam leaderboard (Oct 2026): AI model scores

As of our last check on 22 Oct 2026 (the source shows no leaderboard date), the highest reported Google score on Humanity's Last Exam is gemini-3.1-pro-preview (thinking high) (Google) at 46.44%, #3 on that leaderboard, per Scale AI and Center for AI Safety.

What it measures

2,500 of the toughest, subject-diverse, multi-modal questions, built as a last academic exam for AI.

2,500 expert-written, closed-ended questions across maths, science and the humanities, designed to be hard for frontier models.

What the score means

Accuracy on public questions

Share of public HLE questions answered correctly (±95% confidence interval in the table).

Results from published AI leaderboards. Compare models within each benchmark; scores from different tests aren’t directly comparable. How we collect these results

Scores by model

Run by Scale AI and Center for AI Safety. Leaderboard date not shown. Last checked 22 Oct 2026. Showing top 20 of 55 models listed.

Highlight a provider: Anthropic OpenAI Google Meta Moonshot

Highlighting Google models. Other providers are dimmed. Show all providers

  • Anthropic
  • OpenAI
  • Google
  • Meta
  • Moonshot
Humanity's Last Exam: Calibration error (lower is better) vs AccuracyScatter plot of 18 models. Leader on Calibration error (lower is better): Fable 5.1 (xhigh) at 20%. Google models highlighted. Full values in the table below.20%30%40%50%20%40%60%ACCURACYCALIBRATION ERROR (LOWER IS BETTER) · FLIPPED, BETTER IS UPo3 (high) (April 2025) (OpenAI): Accuracy 20.32%, Calibration error (lower is better) 34%Ogpt-5.1-thinking (OpenAI): Accuracy 23.68%, Calibration error (lower is better) 55%Okimi-k2.5 (Moonshot): Accuracy 24.37%, Calibration error (lower is better) 67%Kclaude-opus-4-5-20251101-thinking (Anthropic): Accuracy 25.20%, Calibration error (lower is better) 55%Agpt-5-2025-08-07 (OpenAI): Accuracy 25.32%, Calibration error (lower is better) 50%Ogpt-5.2-2025-12-11 (OpenAI): Accuracy 27.80%, Calibration error (lower is better) 45%Ogpt-5-pro-2025-10-06 (OpenAI): Accuracy 31.64%, Calibration error (lower is better) 49%Oclaude-opus-4-6-thinking-max (Anthropic): Accuracy 34.44%, Calibration error (lower is better) 46%Aclaude-opus-4-7 (Anthropic): Accuracy 36.20%, Calibration error (lower is better) 47%Agpt-5.4-2026-03-05 (xhigh thinking) (OpenAI): Accuracy 36.24%, Calibration error (lower is better) 42%OMuse Spark (Meta): Accuracy 40.56%, Calibration error (lower is better) 50%Mgpt-5.4-pro-2026-03-05 (OpenAI): Accuracy 44.32%, Calibration error (lower is better) 38%OFable 5.1 (xhigh) (Anthropic): Accuracy 46.50%, Calibration error (lower is better) 20%AGPT 6 Astra (OpenAI): Accuracy 54.80%, Calibration error (lower is better) 39%Ogemini-2.5-pro-preview-06-05 (Google): Accuracy 21.64%, Calibration error (lower is better) 72%Ggemini-3-pro-preview (Google): Accuracy 37.52%, Calibration error (lower is better) 57%GGemini 3.8 Flash (Google): Accuracy 44.52%, Calibration error (lower is better) 51%Ggemini-3.1-pro-preview (thinking high) (Google): Accuracy 46.44%, Calibration error (lower is better) 51%Ggemini-3.1-pro-previewGemini 3.8 FlashFable 5.1
Humanity's Last Exam: Calibration error (lower is better) vs AccuracyScatter plot of 18 models. Leader on Calibration error (lower is better): Fable 5.1 (xhigh) at 20%. Google models highlighted. Full values in the table below.20%30%40%50%20%40%60%ACCURACYCALIBRATION ERROR (LOWER IS BETTER) · FLIPPED, BETTER IS UPo3 (high) (April 2025) (OpenAI): Accuracy 20.32%, Calibration error (lower is better) 34%Ogpt-5.1-thinking (OpenAI): Accuracy 23.68%, Calibration error (lower is better) 55%Okimi-k2.5 (Moonshot): Accuracy 24.37%, Calibration error (lower is better) 67%Kclaude-opus-4-5-20251101-thinking (Anthropic): Accuracy 25.20%, Calibration error (lower is better) 55%Agpt-5-2025-08-07 (OpenAI): Accuracy 25.32%, Calibration error (lower is better) 50%Ogpt-5.2-2025-12-11 (OpenAI): Accuracy 27.80%, Calibration error (lower is better) 45%Ogpt-5-pro-2025-10-06 (OpenAI): Accuracy 31.64%, Calibration error (lower is better) 49%Oclaude-opus-4-6-thinking-max (Anthropic): Accuracy 34.44%, Calibration error (lower is better) 46%Aclaude-opus-4-7 (Anthropic): Accuracy 36.20%, Calibration error (lower is better) 47%Agpt-5.4-2026-03-05 (xhigh thinking) (OpenAI): Accuracy 36.24%, Calibration error (lower is better) 42%OMuse Spark (Meta): Accuracy 40.56%, Calibration error (lower is better) 50%Mgpt-5.4-pro-2026-03-05 (OpenAI): Accuracy 44.32%, Calibration error (lower is better) 38%OFable 5.1 (xhigh) (Anthropic): Accuracy 46.50%, Calibration error (lower is better) 20%AGPT 6 Astra (OpenAI): Accuracy 54.80%, Calibration error (lower is better) 39%Ogemini-2.5-pro-preview-06-05 (Google): Accuracy 21.64%, Calibration error (lower is better) 72%Ggemini-3-pro-preview (Google): Accuracy 37.52%, Calibration error (lower is better) 57%GGemini 3.8 Flash (Google): Accuracy 44.52%, Calibration error (lower is better) 51%Ggemini-3.1-pro-preview (thinking high) (Google): Accuracy 46.44%, Calibration error (lower is better) 51%Ggemini-3.1-pro-previewGemini 3.8 FlashFable 5.1
Humanity's Last Exam: top 15 by AccuracyBars from zero. Leader: GPT 6 Astra at 54.80%. Google models highlighted. Full values in the table below.0%20%40%60%GPT 6 Astra (OpenAI), rank 1: 54.80%OGPT 6 Astra54.80%Fable 5.1 (xhigh) (Anthropic), rank 2: 46.50%AFable 5.146.50%gemini-3.1-pro-preview (thinking high) (Google), rank 3: 46.44%Ggemini-3.1-pro-preview46.44%Sonnet 5.5 (xhigh) (Anthropic), rank 4: 46.12%ASonnet 5.546.12%Gemini 3.8 Flash (Google), rank 5: 44.52%GGemini 3.8 Flash44.52%gpt-5.4-pro-2026-03-05 (OpenAI), rank 6: 44.32%Ogpt-5.4-pro-2026-03-0544.32%Muse Spark (Meta), rank 7: 40.56%MMuse Spark40.56%gemini-3-pro-preview (Google), rank 8: 37.52%Ggemini-3-pro-preview37.52%gpt-5.4-2026-03-05 (xhigh thinking) (OpenAI), rank 9: 36.24%Ogpt-5.4-2026-03-0536.24%claude-opus-4-7 (Anthropic), rank 10: 36.20%Aclaude-opus-4-736.20%claude-opus-4-6-thinking-max (Anthropic), rank 11: 34.44%Aclaude-opus-4-6-thinking-max34.44%gpt-5-pro-2025-10-06 (OpenAI), rank 12: 31.64%Ogpt-5-pro-2025-10-0631.64%Claude Haiku 5.5 (Anthropic), rank 13: 30.12%AClaude Haiku 5.530.12%gpt-5.2-2025-12-11 (OpenAI), rank 14: 27.80%Ogpt-5.2-2025-12-1127.80%gpt-5-2025-08-07 (OpenAI), rank 15: 25.32%Ogpt-5-2025-08-0725.32%gemini-2.5-pro-preview-06-05 (Google), rank 19: 21.64%G#19 gemini-2.5-pro-preview-…21.64%
Humanity's Last Exam: top 12 by AccuracyBars from zero. Leader: GPT 6 Astra at 54.80%. Google models highlighted. Full values in the table below.0%20%40%60%GPT 6 Astra (OpenAI), rank 1: 54.80%OGPT 6 Astra54.80%Fable 5.1 (xhigh) (Anthropic), rank 2: 46.50%AFable 5.146.50%gemini-3.1-pro-preview (thinking high) (Google), rank 3: 46.44%Ggemini-3.1-pro-p…46.44%Sonnet 5.5 (xhigh) (Anthropic), rank 4: 46.12%ASonnet 5.546.12%Gemini 3.8 Flash (Google), rank 5: 44.52%GGemini 3.8 Flash44.52%gpt-5.4-pro-2026-03-05 (OpenAI), rank 6: 44.32%Ogpt-5.4-pro-2026…44.32%Muse Spark (Meta), rank 7: 40.56%MMuse Spark40.56%gemini-3-pro-preview (Google), rank 8: 37.52%Ggemini-3-pro-pre…37.52%gpt-5.4-2026-03-05 (xhigh thinking) (OpenAI), rank 9: 36.24%Ogpt-5.4-2026-03-…36.24%claude-opus-4-7 (Anthropic), rank 10: 36.20%Aclaude-opus-4-736.20%claude-opus-4-6-thinking-max (Anthropic), rank 11: 34.44%Aclaude-opus-4-6-…34.44%gpt-5-pro-2025-10-06 (OpenAI), rank 12: 31.64%Ogpt-5-pro-2025-1…31.64%gemini-2.5-pro-preview-06-05 (Google), rank 19: 21.64%G#19 gemini-2.5-p…21.64%
Accuracy
Share of public HLE questions answered correctly (±95% confidence interval in the table).
Calibration error (lower is better)
RMS gap between the model’s stated confidence and its actual accuracy. High means overconfident.

Model names as listed by Scale. Two models have no published calibration error, so they appear in the table and ranked view only.

All scores

Humanity's Last Exam scores, Google models highlighted
RankModelProviderAccuracyCalibration error (lower is better)±
1 GPT 6 AstraOpenAI54.80%39%±1.94
2 Fable 5.1 (xhigh)Anthropic46.50%20%±2
3 gemini-3.1-pro-preview (thinking high)Google46.44%51%±1.96
4 Sonnet 5.5 (xhigh)Anthropic46.12%–±1.95
5 Gemini 3.8 FlashGoogle44.52%51%±1.96
6 gpt-5.4-pro-2026-03-05OpenAI44.32%38%±1.95
7 Muse SparkMeta40.56%50%±1.92
8 gemini-3-pro-previewGoogle37.52%57%±1.9
9 gpt-5.4-2026-03-05 (xhigh thinking)OpenAI36.24%42%±1.88
10 claude-opus-4-7Anthropic36.20%47%±1.88
Show 10 more rows
RankModelProviderAccuracyCalibration error (lower is better)±
11 claude-opus-4-6-thinking-maxAnthropic34.44%46%±1.86
12 gpt-5-pro-2025-10-06OpenAI31.64%49%±1.82
13 Claude Haiku 5.5Anthropic30.12%–±1.83
14 gpt-5.2-2025-12-11OpenAI27.80%45%±1.76
15 gpt-5-2025-08-07OpenAI25.32%50%±1.7
16 claude-opus-4-5-20251101-thinkingAnthropic25.20%55%±1.7
17 kimi-k2.5Moonshot24.37%67%±1.81
18 gpt-5.1-thinkingOpenAI23.68%55%±1.67
19 gemini-2.5-pro-preview-06-05Google21.64%72%±1.61
20 o3 (high) (April 2025)OpenAI20.32%34%±1.58

Source: Scale AI and Center for AI Safety leaderboard ↗

Best-ranked Google model here: gemini-3.1-pro-preview (thinking high) (Google). Try prompts for Gemini →