ARC-AGI

ARC-AGI — the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.

reasoning

Models Tested

98.0

Top Score

39.2

Average Score

Rankings

#	Model	Score
1	Gemini 3.1 Pro PreviewGoogle DeepMind	98.0
2	Claude Opus 4.6Anthropic	94.0
3	GPT-5.2 ChatOpenAI	86.2
4	GPT-5.2OpenAI	86.2
5	Claude Opus 4.5Anthropic	80.0
6	Gemini 3 ProGoogle	75.0
7	GPT-5.1OpenAI	72.8
8	GPT-5.1 ChatOpenAI	72.8
9	GPT-5 ProOpenAI	70.2
10	GPT-5 ChatOpenAI	65.7
11	GPT-5OpenAI	65.7
12	o3OpenAI	60.8
13	o3 ProOpenAI	59.3
14	o4 MiniOpenAI	58.7
15	Claude Sonnet 4Anthropic	40.0
16	Claude Opus 4Anthropic	35.7
17	o3 MiniOpenAI	34.5
18	Gemini 2.5 FlashGoogle DeepMind	32.3
19	Gemini 2.5 ProGoogle DeepMind	31.3
20	o1OpenAI	30.7
21	Claude 3.7 SonnetAnthropic	28.6
22	Claude 3.7 Sonnet (thinking)Anthropic	28.6
23	DeepSeek-R1 (May 2025)DeepSeek	21.2
24	o1-previewOpenAI	18.0
25	Grok 3 MinixAI	16.5
26	Grok 3 Mini BetaxAI	16.5
27	DeepSeek-R1DeepSeek	15.8
28	o1-miniOpenAI	14.0
29	GPT-4.5OpenAI	10.3
30	Grok 3xAI	5.5
31	GPT-4.1OpenAI	5.5
32	Grok 3 BetaxAI	5.5
33	GPT-4o (2024-11-20)OpenAI	4.5
34	Llama 4 MaverickMeta	4.4
35	GPT-4.1 MiniOpenAI	3.5
36	Llama 4 ScoutMeta	0.5
37	GPT-4.1 NanoOpenAI	0.1