Skip to content

Abstraction and Reasoning Corpus for AGI v2

ARC-AGI-2 tests whether a system can infer a transformation from example grids and apply it to a new grid. Compare accuracy with the evaluation split, attempt limit, and cost attached.

41 models

Higher is better.

  1. 1GPT-6 AstraOpenAI95.0
  2. 2GPT-6.1 SolOpenAI94.2
  3. 3GPT-5.6 SolOpenAI92.5
  4. 4Claude Opus 5.5Anthropic91.7
  5. 5Gemini 4 ArgonGoogle91.3
  6. 6Claude Opus 5Anthropic90.4
  7. 7Claude Fable 5.1Anthropic90.0
  8. 8GPT-6 SolOpenAI89.6
  9. 9Claude Fable 5Anthropic89.2
  10. 10Gemini 3.8 FlashGoogle89.2
  11. 11GPT-5.5OpenAI85.0
  12. 12Gemini 3.7 FlashGoogle84.6
  13. 13GPT-5.5 ProOpenAI84.2
  14. 14GPT-5.6 TerraOpenAI83.9
  15. 15GPT-5.4 ProOpenAI83.3
  16. 16dots3-note PreviewDots Studio81.4
  17. 17Gemini 3.1 ProGoogle77.1
  18. 18Claude Opus 4.7 (Adaptive)Anthropic75.8
  19. 19GPT-5.4OpenAI74.0
  20. 20Gemini 3.5 FlashGoogle72.1
  21. 21Claude Opus 4.8Anthropic72.1
  22. 22Grok 4.6xAI67.1
  23. 23DeepSeek V4 Flash 0731DeepSeek61.4
  24. 24Grok 4.7xAI61.4
  25. 25Claude Haiku 5.5Anthropic61.3
  26. 26DeepSeek V4 Pro 0813DeepSeek61.3
  27. 27Kimi K3Moonshot AI60.4
  28. 28Gemini 3.6 FlashGoogle60.4
  29. 29GPT-5.6 LunaOpenAI59.5
  30. 30GPT-6 LunaOpenAI59.3
  31. 31Claude Sonnet 4.6Anthropic58.3
  32. 32Grok 4.20xAI53.3
  33. 33GPT-5.2OpenAI52.9
  34. 34Grok 4.5xAI52.6
  35. 35Muse SparkMeta42.5
  36. 36Inkling-SmallThinking Machines Lab40.1
  37. 37Claude Opus 4.5 ThinkingAnthropic37.6
  38. 38Gemini 3 ProGoogle31.1
  39. 39GPT-5.4 miniOpenAI18.9
  40. 40Claude Sonnet 4.5Anthropic13.6
Higher is better. Showing the top 40 of 41. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Grid transformation puzzles with novel rules
Tasks
Visual pattern completion and abstract reasoning
Authors
Francois Chollet, ARC Prize Foundation (2025)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.

All results

1GPT-6 Astra OpenAI95.0
2GPT-6.1 Sol OpenAI94.2
3GPT-5.6 Sol OpenAI92.5
4Claude Opus 5.5 Anthropic91.7
5Gemini 4 Argon Google91.3
6Claude Opus 5 Anthropic90.4
7Claude Fable 5.1 Anthropic90.0
8GPT-6 Sol OpenAI89.6
9Claude Fable 5 Anthropic89.2
10Gemini 3.8 Flash Google89.2
11GPT-5.5 OpenAI85.0
12Gemini 3.7 Flash Google84.6
13GPT-5.5 Pro OpenAI84.2
14GPT-5.6 Terra OpenAI83.9
15GPT-5.4 Pro OpenAI83.3
16dots3-note Preview Dots Studio81.4
17Gemini 3.1 Pro Google77.1
18Claude Opus 4.7 (Adaptive) Anthropic75.8
19GPT-5.4 OpenAI74.0
20Gemini 3.5 Flash Google72.1
21Claude Opus 4.8 Anthropic72.1
22Grok 4.6 xAI67.1
23DeepSeek V4 Flash 0731 DeepSeek61.4
24Grok 4.7 xAI61.4
25Claude Haiku 5.5 Anthropic61.3
26DeepSeek V4 Pro 0813 DeepSeek61.3
27Kimi K3 Moonshot AI60.4
28Gemini 3.6 Flash Google60.4
29GPT-5.6 Luna OpenAI59.5
30GPT-6 Luna OpenAI59.3
31Claude Sonnet 4.6 Anthropic58.3
32Grok 4.20 xAI53.3
33GPT-5.2 OpenAI52.9
34Grok 4.5 xAI52.6
35Muse Spark Meta42.5
36Inkling-Small Thinking Machines Lab40.1
37Claude Opus 4.5 Thinking Anthropic37.6
38Gemini 3 Pro Google31.1
39GPT-5.4 mini OpenAI18.9
40Claude Sonnet 4.5 Anthropic13.6
41GPT-5.4 nano OpenAI5.7