Skip to content
PromptLab by FEHMER TECH
ENDE

AI ranking: Humanity's Last Exam (HLE), live

How good are the best AI models on extremely hard expert questions? The chart and leaderboard come live from the benchmark authors' dashboard (Center for AI Safety and Scale AI) and are refreshed hourly.

LiveLast updated: · Data: Center for AI Safety / Scale AI (dashboard.safe.ai, agi.safe.ai)

Top model
54.6 %
Fable 5.1 · Anthropic
Models above 50%
4
of 60
Newest model in the ranking
Muse Spark 1.3
4 September 2026
Models in the ranking
60
13 record-setters

AI progress on HLE

Each dot is a model (release date against accuracy). The line follows the frontier: the best standard model so far.

0%10%20%30%40%50%60%Jan 25Apr 25Jul 25Oct 25Jan 26Apr 26Jul 26GPT-4o (OpenAI): 2.7 %, 20/11/2024o1 (OpenAI): 8.0 %, 05/12/2024DeepSeek R1 (DeepSeek): 8.5 %, 20/01/2025o3-mini (OpenAI): 13.4 %, 31/01/2025Sonnet 3.7 (Anthropic): 8.0 %, 24/02/2025Gemini 2.5 Pro-Exp (Google): 18.8 %, 25/03/2025o3 (OpenAI): 20.3 %, 16/04/2025o4-mini (OpenAI): 18.1 %, 16/04/2025Sonnet 4 (Anthropic): 9.6 %, 22/05/2025Gemini 2.5 Pro (Google): 21.6 %, 17/06/2025Gemini 2.5 Flash (Google): 12.1 %, 17/06/2025Gemini 2.5 Flash-Lite (Google): 6.7 %, 17/06/2025Grok 4 (xAI): 24.5 %, 09/07/2025GPT-5 (OpenAI): 25.3 %, 07/08/2025GPT-5-mini (OpenAI): 19.4 %, 07/08/2025GPT-5-Nano (OpenAI): 8.2 %, 07/08/2025Grok 4 Fast (xAI): 17.8 %, 19/09/2025Sonnet 4.5 (Anthropic): 14.7 %, 29/09/2025Haiku 4.5 (Anthropic): 9.7 %, 15/10/2025Kimi K2 (Moonshot): 21.4 %, 06/11/2025Gemini 3 Pro (Google): 38.3 %, 12/11/2025GPT-5.1 (OpenAI): 27.2 %, 12/11/2025Grok 4.1 Fast (xAI): 18.4 %, 19/11/2025Opus 4.5 (Anthropic): 25.8 %, 24/11/2025DeepSeek 3.2 (DeepSeek): 21.8 %, 01/12/2025GPT-5.2 (OpenAI): 29.9 %, 11/12/2025Gemini 3 Flash (Google): 36.6 %, 17/12/2025Kimi K2.5 (Moonshot): 25.4 %, 29/01/2026Opus 4.6 (Anthropic): 34.2 %, 05/02/2026Sonnet 4.6 (Anthropic): 21.1 %, 17/02/2026Gemini 3.1 Pro (Google): 45.9 %, 19/02/2026GPT-5.4 (OpenAI): 40.3 %, 03/03/2026Gemini 3.1 Flash-Lite (Google): 17.4 %, 03/03/2026Grok 4.2 (xAI): 30.2 %, 09/03/2026GPT-5.4-mini (OpenAI): 23.5 %, 17/03/2026GPT-5.4-Nano (OpenAI): 18.6 %, 17/03/2026GLM 5.1 (Z.ai (Zhipu)): 25.6 %, 06/04/2026Opus 4.7 (Anthropic): 39.0 %, 16/04/2026Grok 4.3 (xAI): 33.1 %, 17/04/2026Kimi K2.6 (Moonshot): 35.5 %, 20/04/2026GPT-5.5 (OpenAI): 43.8 %, 23/04/2026DeepSeek 4 Pro (DeepSeek): 32.4 %, 24/04/2026Gemini 3.5 Flash (Google): 42.5 %, 20/05/2026Opus 4.8 (Anthropic): 42.2 %, 30/05/2026Fable 5 (Anthropic): 52.7 %, 09/06/2026GLM 5.2 (Z.ai (Zhipu)): 41.4 %, 16/06/2026Grok 4.5 (xAI): 39.3 %, 08/07/2026Muse Spark 1.1 (Meta (Muse)): 49.4 %, 09/07/2026GPT-5.6 Sol (OpenAI): 45.5 %, 09/07/2026GPT-5.6 Terra (OpenAI): 35.9 %, 09/07/2026GPT-5.6 Luna (OpenAI): 31.8 %, 09/07/2026Kimi K3 (Moonshot): 41.1 %, 16/07/2026Opus 5 (Anthropic): 51.0 %, 24/07/2026Gemini 3.7 Flash (Google): 45.6 %, 01/08/2026Grok 4.6 (xAI): 39.7 %, 12/08/2026GLM 5.3 (Z.ai (Zhipu)): 39.1 %, 14/08/2026Fable 5.1 (Anthropic): 54.6 %, 01/09/2026Gemini 3.8 Flash (Google): 46.2 %, 02/09/2026GPT-6 Astra (OpenAI): 53.6 %, 03/09/2026Muse Spark 1.3 (Meta (Muse)): 43.2 %, 04/09/2026GPT-4oDeepSeek R1Gemini 2.5 Pro-ExpGemini 2.5 ProGPT-5Gemini 3.1 ProFable 5Fable 5.1
AnthropicOpenAIMeta (Muse)GoogleZ.ai (Zhipu)MoonshotxAIDeepSeek

● Standard · ○ Mini / Nano

Leaderboard (HLE accuracy)

#ModelProviderHLEReleasedSize
1Fable 5.1Anthropic54.6 %Standard
2GPT-6 AstraOpenAI53.6 %Standard
3Fable 5Anthropic52.7 %Standard
4Opus 5Anthropic51.0 %Standard
5Muse Spark 1.1Meta (Muse)49.4 %Standard
6Gemini 3.8 FlashGoogle46.2 %Mini
7Gemini 3.1 ProGoogle45.9 %Standard
8Gemini 3.7 FlashGoogle45.6 %Mini
9GPT-5.6 SolOpenAI45.5 %Standard
10GPT-5.5OpenAI43.8 %Standard
11Muse Spark 1.3Meta (Muse)43.2 %Standard
12Gemini 3.5 FlashGoogle42.5 %Mini
13Opus 4.8Anthropic42.2 %Standard
14GLM 5.2Z.ai (Zhipu)41.4 %Standard
15Kimi K3Moonshot41.1 %Standard
16GPT-5.4OpenAI40.3 %Standard
17Grok 4.6xAI39.7 %Standard
18Grok 4.5xAI39.3 %Standard
19GLM 5.3Z.ai (Zhipu)39.1 %Standard
20Opus 4.7Anthropic39.0 %Standard
21Gemini 3 ProGoogle38.3 %Standard
22Gemini 3 FlashGoogle36.6 %Mini
23GPT-5.6 TerraOpenAI35.9 %Mini
24Kimi K2.6Moonshot35.5 %Standard
25Opus 4.6Anthropic34.2 %Standard
26Grok 4.3xAI33.1 %Standard
27DeepSeek 4 ProDeepSeek32.4 %Standard
28GPT-5.6 LunaOpenAI31.8 %Mini
29Grok 4.2xAI30.2 %Standard
30GPT-5.2OpenAI29.9 %Standard

More benchmarks

The same source provides more tests. Not every model was run on every test.

ARC-AGI-2

  1. 1.GPT-5.577.5 %
  2. 2.Gemini 3.1 Pro73.3 %
  3. 3.Opus 4.867.5 %
  4. 4.GPT-5.465.0 %
  5. 5.Gemini 3.5 Flash63.3 %
  6. 6.Grok 4.255.0 %
  7. 7.Opus 4.750.8 %
  8. 8.Opus 4.644.3 %

SWE-bench Pro

  1. 1.Fable 577.4 %
  2. 2.Opus 4.865.4 %
  3. 3.Opus 4.760.9 %
  4. 4.Grok 4.560.3 %
  5. 5.GPT-5.6 Sol57.9 %
  6. 6.Opus 4.656.7 %
  7. 7.Muse Spark 1.154.3 %
  8. 8.Gemini 3.5 Flash53.9 %

HLE Diamond

  1. 1.GPT-6 Astra59.9 %
  2. 2.Fable 5.150.7 %
  3. 3.Fable 545.6 %
  4. 4.Opus 538.1 %
  5. 5.Gemini 3.8 Flash33.3 %
  6. 6.Gemini 3.1 Pro31.8 %
  7. 7.Muse Spark 1.131.3 %
  8. 8.Gemini 3.7 Flash31.2 %

What is Humanity's Last Exam?

HLE is a benchmark of 2,500 very hard expert questions, built by the Center for AI Safety and Scale AI with subject experts worldwide. It measures how close AI models are to human expert knowledge now that older tests like MMLU are saturated.

agi.safe.aidashboard.safe.aiarXiv 2501.14249

How do I read the values?

The number is the share of questions answered correctly. Results differ by method: with or without tools, with or without images. This page shows the dashboard's values unchanged; vendor-reported numbers may differ.

FAQ about the HLE ranking

What is Humanity's Last Exam (HLE)?

HLE is a benchmark of 2,500 very hard expert questions, built by the Center for AI Safety and Scale AI with subject experts worldwide. It shows how close AI models come to human expert knowledge.

Which AI model currently leads the HLE ranking?

According to the benchmark authors' dashboard, Fable 5.1 by Anthropic currently leads with 54.6% (data as of 12 October 2026). 4 models are above 50%.

Why do HLE scores differ between sources?

The method differs: with or without tools (web search, code), with or without image questions, different grading models. This page shows the dashboard's values unchanged; vendor-reported numbers can be higher.

How often is the ranking updated?

The ranking is loaded live from the dashboard every hour. If the source is down, the page shows the last stored copy and labels it.