ARC-AGI 2
11 models tested · Updated 2026-02-05 · Verified sources only
Claude Opus 4.6 leads at 85.0%
1
Anthropic · Blog/Anthropic · 2026-02-05
Max effort config with 120k thinking budget. Standard config scored 68.8%.
85.0%
2
OpenAI · OpenAI Blog · 2026-04-23
Huge 12pt jump over GPT-5.4 (73.3). Still below Gemini 3.1 Pro (77.1).
85.0%
3
Google · Blog/Google · 2026-02-12
ARC Prize Foundation verified. Specialized reasoning mode — highest ARC-AGI-2 score by a wide margin (prev best ~54%). Also scored Codeforces Elo 3455.
84.6%
4
OpenAI · Blog/OpenAI · 2026-03-05
Pro variant. Near Gemini 3 Deep Think (84.6%). Highest commercial model score.
83.3%
5
Google · Blog/Google · 2026-02-19
Verified from Google blog. More than 2x Gemini 3 Pro reasoning performance.
77.1%
6
Anthropic · Blog/OpenAI · 2026-04-16
OpenAI-tested. Lags GPT-5.5 by ~10pp on ARC-AGI-2.
75.8%
7
OpenAI · Blog/OpenAI · 2026-03-05
Up from 52.9% on GPT-5.2. Large jump in abstract reasoning.
73.3%
8
Google DeepMind · Blog/Google DeepMind · 2026-06-10
Below GPT-5.5 (84.6) and Gemini 3.1 Pro (77.1). Up from Gemini 3 Flash (33.6).
72.1%
9
Meta · Blog/Meta · 2026-04-08
Below frontier models (Claude Opus 4.6 at 85%, GPT-5.4 at 73.3%).
42.5%
10
Google DeepMind · Blog/Google DeepMind · 2026-06-10
Below Gemini 3.5 Flash (72.1). Big gain for new Flash.
33.6%
11
xAI · Blog/xAI · 2025-07-09
Nearly double Opus ~8.6%. New SOTA for closed models on ARC-AGI v2.
15.9%