benchmarks.site

Public leaderboard · General knowledge & reasoning

ARC-AGI (ARC Prize)

Puzzles that most people find easy but that require spotting a brand-new pattern rather than recalling facts; the newest version (ARC-AGI-3) drops the AI into small interactive games it has never seen. It is a popular test of whether AI can genuinely learn new skills on the fly.

Leaderboard

Top 3 right now

  1. 1stGPT-6 Astra - Provider Adapter (High)99.9
  2. 2ndGemini 4 Carbon - Provider Adapter (Medium)99.1
  3. 3rdGPT-6 Astra - Provider Adapter (Max)98.6

What it measures

Novel abstraction puzzles that are easy for humans (grid transformations in ARC-AGI-1 and 2, interactive game environments in ARC-AGI-3), scored as the share of tasks solved and plotted against cost per task.

How to read the numbers

Share of ARC-AGI-3 game environments solved on the semi-private test set (%) — higher is better.

Results

25 rows · higher is better (ARC-AGI-3 score (%))CSV
1GPT-6 Astra - Provider Adapter (High)OpenAI99.918816.63
2Gemini 4 Carbon - Provider Adapter (Medium)Google99.14559.50
3GPT-6 Astra - Provider Adapter (Max)OpenAI98.617331.95
4GPT-6 Astra - Provider Adapter (XHigh)OpenAI98.418147.19
5GPT-6 Astra - Provider Adapter (Medium)OpenAI98.419284.78
6GPT-6 Astra - Provider Adapter (Low)OpenAI98.021297.92
7Claude Fable 5.2 - Provider Adapter (High)Anthropic97.718624.91
8Gemini 4 Carbon - Provider Adapter (Low)Google97.74350.87
9Gemini 4 Argon - Provider Adapter (Low)Google97.54770.15
10Gemini 4 Argon - Provider Adapter (Medium)Google97.34981.20
11Gemini 4 Argon - Provider Adapter (High)Google97.34625.84
12Claude Fable 5.2 - Provider Adapter (XHigh)Anthropic97.318251.68
13Gemini 4 Carbon - Provider Adapter (High)Google96.94350.03
14Claude Fable 5.2 - Provider Adapter (Medium)Anthropic96.821662.29
15GPT-6 Astra - Provider Adapter (None)OpenAI96.723456.66
16GPT-6.1 Sol - Provider Adapter (XHigh)OpenAI96.44360.21
17GPT-6.1 Sol - Provider Adapter (Max)OpenAI96.23816.80
18GPT-6.1 Sol - Provider Adapter (High)OpenAI95.05403.98
19Claude Fable 5.2 - Provider Adapter (Max)Anthropic94.917875.84
20Claude Fable 5.2 - Provider Adapter (Low)Anthropic92.327377.33
21GPT-6.1 Sol - Provider Adapter (Medium)OpenAI91.05834.91
22GPT-6.1 Sol - Provider Adapter (Low)OpenAI82.86622.20
23Claude Fable 5.2 (XHigh)Anthropic81.221948.83
24Claude Fable 5.2 (High)Anthropic79.325173.94
25Claude Fable 5.2 (Medium)Anthropic73.329034.48

Note: ARC-AGI-3 semi-private set, from the leaderboard data file behind arcprize.org/leaderboard; raw fractions converted to % (1 decimal). Rows are system/effort configurations, several per model; 'Provider Adapter' entries use a provider-specific harness. Cost is as listed in the file (units per ARC Prize; shown to 2 decimals). ARC-AGI-1 and ARC-AGI-2 are near-saturated (top systems ~98.5% and ~95%) and are also in v1.json/v2.json.

Source: ARC Prize Foundation · arcprize.org · results as of 2026-11-16 · retrieved 2026-11-17

Markdown · More General knowledge & reasoning