ARC-AGI-2 what it measures, how it is scored and who leads it

Name: ARC-AGI-2
Creator: ARC Prize Foundation
Published: 2025-03-24

Reasoning · ARC Prize Foundation · introduced Mar 24, 2025 · Tasks solved

Leaderboard Paper

Scores from ARC-AGI-2, published Sep 24, 2026

Rank	Model	Tasks solved
1	GPT-6 AstraOpenAI · as “GPT-6 Astra (Max)”	95.0%
2	Claude Opus 5.5Anthropic · as “Claude Opus 5.5 (High)”	93.3%
3	GPT-5.6 SolOpenAI · as “GPT-5.6 Sol (Max)”	92.5%
4	Claude Opus 5Anthropic · as “Claude Opus 5 (Max)”	90.4%
5	Claude Fable 5.1Anthropic · as “Claude Fable 5.1 (Max)”	90.0%
6	Claude Fable 5Anthropic · as “Claude Fable 5 (Max)”	89.2%
6	Gemini 3.8 FlashGoogle · as “Gemini 3.8 Flash (High)”	89.2%
8	GPT-5.5OpenAI · as “GPT-5.5 (XHigh)”	85.0%
9	Gemini 3.7 FlashGoogle · as “Gemini 3.7 Flash (High)”	84.6%
10	Gemini 3Google · as “Gemini 3 Deep Think (2/26)”	84.6%
10	GPT-5.5 ProOpenAI · as “GPT-5.5 Pro (High)”	84.6%
12	GPT-5.6 TerraOpenAI · as “GPT-5.6 Terra (Max)”	83.9%
13	GPT-5.4 ProOpenAI · as “GPT-5.4 Pro (XHigh)”	83.3%
14	Gemini 3.1 Pro PreviewGoogle · as “Gemini 3.1 Pro (Preview)”	77.1%
15	Dots3 NoteDots Studio · as “Dots3-Note Preview (Max)”	76.8%
16	Claude 4.7Anthropic · as “Claude 4.7 (Max)”	75.8%
17	GPT-5.4OpenAI · as “GPT-5.4 (XHigh)”	74.0%
18	Claude Opus 4.8Anthropic · as “Claude Opus 4.8 (High)”	72.1%
18	Gemini 3.5 FlashGoogle · as “Gemini 3.5 Flash (High)”	72.1%
20	Claude Opus 4.6Anthropic · as “Claude Opus 4.6 (120K, High)”	69.2%
21	Grok 4.6SpaceXAI · as “Grok 4.6 (XHigh)”	67.1%
22	Grok 4.20SpaceXAI · as “Grok 4.20 (Reasoning)”	65.1%
23	DeepSeek V4 Flash VisionDeepSeek · as “DeepSeek V4 Flash 0731 (Max)”	61.4%
24	DeepSeek V4 Pro 0813DeepSeek · as “DeepSeek V4 Pro 0813 (Max)”	61.3%
25	Claude Sonnet 4.6Anthropic · as “Claude Sonnet 4.6 (High)”	60.4%

About this benchmark

Task: Solve harder abstract grid puzzles that are easy for people and hard for AI.
Dataset: 1,000 public training tasks; public, semi-private and private evaluation sets of 120 tasks each, every task solved by at least two people.
Method: Two attempts per task (pass@2), cost per task reported next to the score.
Metric: Tasks solved
Organization: ARC Prize Foundation
Introduced: Mar 24, 2025
Paper: ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems (arXiv 2505.11831)
Official leaderboard: arcprize.org/leaderboard
Home: arcprize.org/arc-agi/2/

Versions

Newest first. New versions are added, never rewritten.

Version	Date	What changed
Technical reportPaper with the human testing baseline.	May 17, 20251 year ago	Paper with the human testing baseline.
2Announced with ARC Prize 2025.	Mar 24, 20251 year ago	Announced with ARC Prize 2025.

Sources: each benchmark's paper (arXiv) and its official site or leaderboard. Scores are the boards captured here every day, exactly as published.

Scores from ARC-AGI-2, published Sep 24, 2026

About this benchmark

Versions

New #1s by email