AI Model Release Tracker
Abstract reasoning

ARC-AGI-2

Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better.

Rankings

Higher is better
#1
GPT-6 Astravia BenchLM
95%
2
GPT-5.6 Solvia BenchLM
92.5%
3
Claude Opus 5via BenchLM
90.4%
4
Claude Fable 5.1
90%
5
GPT-5.5
84.6%
6
GPT-5.6 Terravia BenchLM
83.9%
7
GPT-5.4-Provia BenchLM
83.3%
8
Gemini 3.1 Pro
77.1%
9
Claude Opus 4.7
75.8%
10
GPT-5.4via BenchLM
73.95%
11
Gemini 3.5 Flash
72.1%
12
Claude Opus 4.8via BenchLM
72.08%
13
GPT-5.6 Lunavia BenchLM
59.54%
14
Claude Sonnet 4.6
58.3%
15
Grok 4.20 Betavia BenchLM
53.3%
16
GPT-5.2via BenchLM
52.9%
17
Grok 4.5via BenchLM
52.64%
18
Muse Sparkvia BenchLM
42.5%
19
Gemini 3.0 Flash
33.6%
20
Gemini 3.0 Provia BenchLM
31.1%
21
Claude Sonnet 4.5via BenchLM
13.6%

Scores marked “via” above are quoted with attribution from BenchLM, retrieved 7 September 2026.

ARC-AGI-2 — frequently asked questions

Puzzle-style tests of abstract reasoning and pattern-finding — the kind of thing people find easy but AIs often struggle with. Higher is better.