Benchmark · ReasoningSettled

ARC-AGI-2

ARC-AGI-2 · the second iteration of the Abstraction and Reasoning Corpus, testing novel pattern recognition and abstract reasoning without prior training data.

Updated 2026-11-17
Models tested
91
Top score
95.8
GPT-6.2 Astra
Median
33.6
min 0.1
Top-5 spread
σ 0.7
Settled

Best score over time · one chart, every benchmark

ARC-AGI-280 MODELS · FRONTIER RUNNING MAX0255075100SCORE ↑Jan 25Jul 25Dec 25Jun 26Nov 26RELEASE DATE →benchgecko.ai/benchmark/arc-agi-2 · frontier
Frontier on ARC-AGI-2 rose from 1.3 to 95.8 in 22 months · +94.5 points · latest leader GPT-6.2 Astra from OpenAI.
Pink dots = frontier records · 19 totalClick to open model page

91 models tested · sorted by score

#ModelScore
1OpenAI logoGPT-6.2 Astra95.8
2OpenAI logoGPT-6 Astra95.0
3Google DeepMind logoGemini 4 Carbon94.2
4OpenAI logoGPT-6.1 Sol94.2
5Anthropic logoClaude Fable 5.293.8
6Anthropic logoClaude Opus 5.593.3
7OpenAI logoGPT-5.6 Sol92.5
8Google DeepMind logoGemini 4 Argon91.3
9Anthropic logoClaude Opus 590.4
10Anthropic logoClaude Fable 5.190.0
11OpenAI logoGPT-6 Sol89.6
12Anthropic logoClaude Fable 589.2
13Google DeepMind logoGemini 3.8 Flash89.2
14OpenAI logoGPT-5.585.0
15Google DeepMind logoGemini 3.7 Flash84.6
16OpenAI logoGPT-5.5 Pro84.6
17OpenAI logoGPT-5.6 Terra83.9
18Anthropic logoClaude Sonnet 5.583.9
19OpenAI logoGPT-5.4 Pro83.3
20DeepSeek logoDeepSeek V4.1 Flash 101778.1
21Google DeepMind logoGemini 3.1 Pro Preview77.1
22Anthropic logoClaude Opus 4.775.8
23OpenAI logoGPT-5.474.0
24DeepSeek logoDeepSeek V4.1 Flash72.9
25Anthropic logoClaude Opus 4.872.1
26Google DeepMind logoGemini 3.5 Flash72.1
27Anthropic logoClaude Opus 4.669.2
28xAI logoGrok 4.667.1
29z-ai logoGLM 5.3 Flash65.8
30xAI logoGrok 4.2065.1
31DeepSeek logoDeepSeek V4 Flash 073161.4
32xAI logoGrok 4.761.4
33Anthropic logoClaude Haiku 5.561.3
34DeepSeek logoDeepSeek V4 Pro 081361.3
35moonshotai logoKimi K360.4
36Google DeepMind logoGemini 3.6 Flash60.4
37Anthropic logoClaude Sonnet 4.660.4
38OpenAI logoGPT-5.6 Luna59.5
39OpenAI logoGPT-6 Luna59.3
40Meta logoMuse Spark Open57.5
41OpenAI logoGPT-5.2 Pro54.2
42OpenAI logoGPT-5.252.9
43xAI logoGrok 4.552.6
44Alibaba Qwen logoQwen3.8 27B42.4
45Anthropic logoClaude Opus 4.537.6
46Google DeepMind logoGemini 3 Flash Preview33.6
47Google DeepMind logoGemini 3 Pro31.1
48z-ai logoGLM 5.222.8
49OpenAI logoGPT-5.4 Mini18.9
50OpenAI logoGPT-5 Pro18.3
51OpenAI logoGPT-5.117.6
52xAI logoGrok 416.0
53Anthropic logoClaude Sonnet 4.513.6
54moonshotai logoKimi K2.511.8
55Google DeepMind logoGemini 3.5 Flash Lite10.3
56OpenAI logoGPT-59.9
57Anthropic logoClaude Opus 48.6
58OpenAI logoo36.5
59OpenAI logoo4 Mini6.1
60OpenAI logoo4 Mini (batch)6.1
61Anthropic logoClaude Sonnet 45.9
62OpenAI logoGPT-5.4 Nano5.7
63xAI logoGrok 4 Fast5.3
64minimax logoMiniMax M2.54.9
65Google DeepMind logoGemini 2.5 Pro4.9
66OpenAI logoo3 Pro4.9
67z-ai logoGLM 54.9
68OpenAI logoGPT-5 Mini4.4
69DeepSeek logoDeepSeek V3.24.0
70Anthropic logoClaude Haiku 4.54.0
71OpenAI logoo3 Mini (batch)3.0
72OpenAI logoo3 Mini3.0
73OpenAI logoGPT-5 Nano2.6
74Google DeepMind logoGemini 2.5 Flash2.5
75DeepSeek logoR11.3
76Google DeepMind logoGemini 2.0 Flash (Dec 2024)1.3
77Google DeepMind logoGemini 2.0 Flash1.3
78Alibaba Qwen logoQwen3 235B A22B Instruct 25071.3
79DeepSeek logoR1 05281.1
80Anthropic logoClaude 3.7 Sonnet0.9
81OpenAI logoo1-mini0.8
82OpenAI logoGPT-4.50.8
83Google DeepMind logoGemini 1.5 Pro (Feb 2024)0.8
84Google DeepMind logoGemini 1.5 Pro (May 2024)0.8
85OpenAI logoGPT-4.10.4
86xAI logoGrok-3 mini0.4
87OpenAI logoGPT-4o-mini (2024-07-18)0.1
88Meta logoLlama 4 Scout0.1
89OpenAI logoGPT-4.1 Mini0.1
90Meta logoLlama 4 Maverick0.1
91OpenAI logoGPT-4.1 Nano0.1

What it measures · how it is scored · sources · as of 2026-11-22

  • What it measures · ARC-AGI 2, harder sequel to ARC. More complex abstract reasoning patterns that test generalization ability beyond training data.
  • How it is scored · Scored as a percentage · higher is better · maximum 100
  • Released · 2025-03-24
  • Scores from · Epoch AI Benchmarking Hub (91 scores)

Same category · related evaluations