Benchmark Almanac

ARC-AGI-2

GPT-6 Astra 95% leads since Sep 2026 In use

Models infer a rule from example grid pairs and produce the output grid for new inputs. Scored by tasks solved within two attempts.

ARC-AGI-3 is its successor, and both have ARC Prize 2026 competitions.

In the authors’ words

The Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI), introduced in 2019, established a challenging benchmark for evaluating the general fluid intelligence of artificial systems via a set of unique, novel tasks only requiring minimal prior knowledge. While ARC-AGI has spurred significant research activity over the past five years, recent AI progress calls for benchmarks capable of finer-grained evaluation at higher levels of cognitive complexity. We introduce ARC-AGI-2, an upgraded version of the benchmark. ARC-AGI-2 preserves the input-output pair task format of its predecessor, ensuring continuity for researchers. It incorporates a newly curated and expanded set of tasks specifically designed to provide a more granular signal to assess abstract reasoning and problem-solving abilities at higher levels of fluid intelligence. To contextualize the difficulty and characteristics of ARC-AGI-2, we present extensive results from human testing, providing a robust baseline that highlights the benchmark's accessibility to human intelligence, yet difficulty for current AI systems. ARC-AGI-2 aims to serve as a next-generation tool for rigorously measuring progress towards more general and human-like AI capabilities.

Abstract of ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems (arXiv:2505.11831, CC0)
Made by ARC Prize Foundation Released Mar 2025 Abstract puzzles and logic Public tasks plus a private test set Last in a launch post Sep 2026 Last model added Oct 2026 Scores refreshed 2 Nov 2026
Level
Everyday Most adults, without training
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Short answer A number, word or short string matched to a key.
arcprize/ARC-AGI-2 746 stars Paper Leaderboard arcprize.org
97
Entries on its own leaderboard, 95 models we track
21
Launch posts list it, from 8 labs
12
System cards report it, from 5 labs

GPT-6 Astra leads with 95%

Best score per entry from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

97 entries. Scroll sideways for the rest.

  1. 95.0 GPT-6 Astra
  2. 94.2 GPT-6.1 Sol
  3. 93.3 Claude Opus 5.5
  4. 92.5 GPT-5.6 Sol
  5. 91.3 Gemini 4 Argon
  6. 90.4 Claude Opus 5
  7. 90.0 Claude Fable 5.1
  8. 89.6 GPT-6 Sol
  9. 89.2 Claude Fable 5
  10. 89.2 Gemini 3.8 Flash
  11. 85.0 GPT-5.5
  12. 84.6 Gemini 3 Deep Think
  13. 84.6 GPT-5.5 Pro
  14. 84.6 Gemini 3.7 Flash
  15. 83.9 GPT-5.6 Terra
  16. 83.9 Claude Sonnet 5.5
  17. 83.3 GPT-5.4 Pro
  18. 78.1 DeepSeek V4.1 Flash 1017
  19. 77.1 Gemini 3.1 Pro
  20. 76.8 Dots3-Note Preview (Max)
  21. 75.8 Claude Opus 4.7
  22. 74.0 GPT-5.4
  23. 72.9 DeepSeek V4.1 Flash
  24. 72.9 GPT-5.2
  25. 72.1 Gemini 3.5 Flash
  26. 72.1 Claude Opus 4.8
  27. 69.2 Claude Opus 4.6
  28. 67.1 Grok 4.6
  29. 65.8 GLM-5.3-Flash
  30. 65.1 Grok 4.20
  31. 61.4 DeepSeek V4 Flash 0731
  32. 61.4 Grok 4.7
  33. 61.3 DeepSeek V4 Pro 0813
  34. 61.3 Claude Haiku 5.5
  35. 60.4 Claude Sonnet 4.6
  36. 60.4 Kimi K3
  37. 60.4 Gemini 3.6 Flash
  38. 59.6 GPT-5.6 Luna
  39. 59.3 GPT-6 Luna
  40. 57.5 Muse Spark Open
  41. 54.2 GPT-5.2 Pro
  42. 54.0 Gemini 3 Pro
  43. 52.6 Grok 4.5
  44. 45.1 Gemini 3 Deep Think (Preview)
  45. 42.4 Qwen 3.8 27B
  46. 40.1 Inkling-Small
  47. 37.6 Claude Opus 4.5
  48. 36.5 Inkling
  49. 33.6 Gemini 3 Flash
  50. 29.4 Grok 4
  51. 23.5 Beam 1.0
  52. 22.8 GLM-5.2
  53. 18.9 GPT-5.4 Mini
  54. 18.3 GPT-5 Pro
  55. 17.6 GPT-5.1
  56. 13.6 Claude Sonnet 4.5
  57. 11.8 Kimi K2.5
  58. 10.3 Gemini 3.5 Flash-Lite
  59. 9.86 GPT-5
  60. 8.61 Claude Opus 4
  61. 6.53 o3
  62. 6.25 tiny-recursion-model
  63. 6.11 o4-mini
  64. 5.93 Claude Sonnet 4
  65. 5.69 GPT-5.4 Nano
  66. 5.28 Grok 4 Fast
  67. 4.86 Gemini 2.5 Pro
  68. 4.86 o3-pro
  69. 4.86 GLM-5
  70. 4.86 MiniMax-M2.5
  71. 4.44 GPT-5 mini
  72. 4.03 Claude Haiku 4.5
  73. 4.03 DeepSeek-V3.2
  74. 2.99 o3-mini
  75. 2.61 GPT-5 nano
  76. 2.54 Gemini 2.5 Flash
  77. 1.94 Nemotron 3.5 Super VL
  78. 1.30 Gemini 2.0 Flash
  79. 1.30 DeepSeek-R1
  80. 1.27 codex-mini-2025-05-16
  81. 1.25 Qwen3-235B-A22B-Instruct
  82. 1.12 DeepSeek-R1
  83. 0.90 Claude 3.7 Sonnet
  84. 0.83 o1-mini
  85. 0.80 Gemini 1.5 Pro
  86. 0.80 GPT-4.5
  87. 0.42 Grok-3 mini
  88. 0.42 GPT-4.1
  89. 0.0 GPT-4o mini
  90. 0.0 GPT-4o
  91. 0.0 Llama 4 Maverick
  92. 0.0 Llama 4 Scout
  93. 0.0 Grok 3
  94. 0.0 GPT-4.1 mini
  95. 0.0 GPT-4.1 nano
  96. 0.0 Magistral Medium 1.0
  97. 0.0 Magistral Small 1.0
  1. 1 GPT-6 Astra max 95%
  2. 2 GPT-6.1 Sol max 94.2%
  3. 3 Claude Opus 5.5 high 93.3%
  4. 4 GPT-5.6 Sol max 92.5%
  5. 5 Gemini 4 Argon high 91.3%
  6. 6 Claude Opus 5 max 90.4%
  7. 7 Claude Fable 5.1 max 90%
  8. 8 GPT-6 Sol max 89.6%
  9. 9 Claude Fable 5 max 89.2%
  10. 10 Gemini 3.8 Flash high 89.2%
  11. 11 GPT-5.5 xhigh 85%
  12. 12 Gemini 3 Deep Think 84.6%
  13. 13 GPT-5.5 Pro high 84.6%
  14. 14 Gemini 3.7 Flash high 84.6%
  15. 15 GPT-5.6 Terra max 83.9%
  16. 16 Claude Sonnet 5.5 xhigh 83.9%
  17. 17 GPT-5.4 Pro xhigh 83.3%
  18. 18 DeepSeek V4.1 Flash 1017 max 78.1%
  19. 19 Gemini 3.1 Pro 77.1%
  20. 20 Dots3-Note Preview (Max) 76.8%
  21. 21 Claude Opus 4.7 max 75.8%
  22. 22 GPT-5.4 xhigh 74%
  23. 23 DeepSeek V4.1 Flash max 72.9%
  24. 24 GPT-5.2 72.9%
  25. 25 Gemini 3.5 Flash high 72.1%
  26. 26 Claude Opus 4.8 high 72.1%
  27. 27 Claude Opus 4.6 120k, high 69.2%
  28. 28 Grok 4.6 xhigh 67.1%
  29. 29 GLM-5.3-Flash max 65.8%
  30. 30 Grok 4.20 65.1%
  31. 31 DeepSeek V4 Flash 0731 max 61.4%
  32. 32 Grok 4.7 xhigh 61.4%
  33. 33 DeepSeek V4 Pro 0813 max 61.3%
  34. 34 Claude Haiku 5.5 xhigh 61.3%
  35. 35 Claude Sonnet 4.6 high 60.4%
  36. 36 Kimi K3 max 60.4%
  37. 37 Gemini 3.6 Flash high 60.4%
  38. 38 GPT-5.6 Luna max 59.6%
  39. 39 GPT-6 Luna max 59.3%
  40. 40 Muse Spark Open xhigh 57.5%
  41. 41 GPT-5.2 Pro high 54.2%
  42. 42 Gemini 3 Pro 54%
  43. 43 Grok 4.5 high 52.6%
  44. 44 Gemini 3 Deep Think (Preview) 45.1%
  45. 45 Qwen 3.8 27B xhigh 42.4%
  46. 46 Inkling-Small xhigh 40.1%
  47. 47 Claude Opus 4.5 thinking, 64k 37.6%
  48. 48 Inkling 36.5%
  49. 49 Gemini 3 Flash high 33.6%
  50. 50 Grok 4 29.4%
  51. 51 Beam 1.0 xhigh 23.5%
  52. 52 GLM-5.2 22.8%
  53. 53 GPT-5.4 Mini xhigh 18.9%
  54. 54 GPT-5 Pro 18.3%
  55. 55 GPT-5.1 thinking, high 17.6%
  56. 56 Claude Sonnet 4.5 thinking 32k 13.6%
  57. 57 Kimi K2.5 11.8%
  58. 58 Gemini 3.5 Flash-Lite high 10.3%
  59. 59 GPT-5 high 9.86%
  60. 60 Claude Opus 4 thinking 16k 8.61%
  61. 61 o3 high 6.53%
  62. 62 tiny-recursion-model 6.25%
  63. 63 o4-mini high 6.11%
  64. 64 Claude Sonnet 4 thinking 16k 5.93%
  65. 65 GPT-5.4 Nano xhigh 5.69%
  66. 66 Grok 4 Fast 5.28%
  67. 67 Gemini 2.5 Pro Jun 2025, thinking 32k 4.86%
  68. 68 o3-pro high 4.86%
  69. 69 GLM-5 4.86%
  70. 70 MiniMax-M2.5 4.86%
  71. 71 GPT-5 mini high 4.44%
  72. 72 Claude Haiku 4.5 thinking 32k 4.03%
  73. 73 DeepSeek-V3.2 4.03%
  74. 74 o3-mini high 2.99%
  75. 75 GPT-5 nano high 2.61%
  76. 76 Gemini 2.5 Flash May 2025, thinking 24k 2.54%
  77. 77 Nemotron 3.5 Super VL medium 1.94%
  78. 78 Gemini 2.0 Flash Dec 2024 1.30%
  79. 79 DeepSeek-R1 1.30%
  80. 80 codex-mini-2025-05-16 1.27%
  81. 81 Qwen3-235B-A22B-Instruct Jul 2025 1.25%
  82. 82 DeepSeek-R1 May 2025 1.12%
  83. 83 Claude 3.7 Sonnet 8k 0.90%
  84. 84 o1-mini 0.83%
  85. 85 Gemini 1.5 Pro Sep 2024 0.80%
  86. 86 GPT-4.5 0.80%
  87. 87 Grok-3 mini low 0.42%
  88. 88 GPT-4.1 0.42%
  89. 89 GPT-4o mini 0%
  90. 90 GPT-4o Nov 2024 0%
  91. 91 Llama 4 Maverick 0%
  92. 92 Llama 4 Scout 0%
  93. 93 Grok 3 0%
  94. 94 GPT-4.1 mini 0%
  95. 95 GPT-4.1 nano 0%
  96. 96 Magistral Medium 1.0 0%
  97. 97 Magistral Small 1.0 0%
#EntryScoreSettingDateSource
1 GPT-6 Astra 95% max Sep 2026 Official leaderboard
2 GPT-6.1 Sol 94.2% max Sep 2026 Official leaderboard
3 Claude Opus 5.5 93.3% high Sep 2026 Official leaderboard
4 GPT-5.6 Sol 92.5% max Jul 2026 Official leaderboard
5 Gemini 4 Argon 91.3% high Sep 2026 Official leaderboard
6 Claude Opus 5 90.4% max Jul 2026 Official leaderboard
7 Claude Fable 5.1 90% max Sep 2026 Official leaderboard
8 GPT-6 Sol 89.6% max Sep 2026 Official leaderboard
9 Claude Fable 5 89.2% max Jun 2026 Official leaderboard
10 Gemini 3.8 Flash 89.2% high Sep 2026 Official leaderboard
11 GPT-5.5 85% xhigh Apr 2026 Official leaderboard
12 Gemini 3 Deep Think 84.6% Feb 2026 Official leaderboard
13 GPT-5.5 Pro 84.6% high Apr 2026 Official leaderboard
14 Gemini 3.7 Flash 84.6% high Aug 2026 Official leaderboard
15 GPT-5.6 Terra 83.9% max Jul 2026 Official leaderboard
16 Claude Sonnet 5.5 83.9% xhigh Sep 2026 Official leaderboard
17 GPT-5.4 Pro 83.3% xhigh Mar 2026 Official leaderboard
18 DeepSeek V4.1 Flash 1017 78.1% max Oct 2026 Official leaderboard
19 Gemini 3.1 Pro 77.1% Feb 2026 Official leaderboard
20 Dots3-Note Preview (Max) 76.8% Aug 2026 Official leaderboard
21 Claude Opus 4.7 75.8% max Apr 2026 Official leaderboard
22 GPT-5.4 74% xhigh Mar 2026 Official leaderboard
23 DeepSeek V4.1 Flash 72.9% max Sep 2026 Official leaderboard
24 GPT-5.2 72.9% Johan Land refinement Dec 2025 Official leaderboard
25 Gemini 3.5 Flash 72.1% high May 2026 Official leaderboard
26 Claude Opus 4.8 72.1% high May 2026 Official leaderboard
27 Claude Opus 4.6 69.2% 120k, high Feb 2026 Official leaderboard
28 Grok 4.6 67.1% xhigh Aug 2026 Official leaderboard
29 GLM-5.3-Flash 65.8% max Aug 2026 Official leaderboard
30 Grok 4.20 65.1% Feb 2026 Official leaderboard
31 DeepSeek V4 Flash 0731 61.4% max Jul 2026 Official leaderboard
32 Grok 4.7 61.4% xhigh Sep 2026 Official leaderboard
33 DeepSeek V4 Pro 0813 61.3% max Aug 2026 Official leaderboard
34 Claude Haiku 5.5 61.3% xhigh Oct 2026 Official leaderboard
35 Claude Sonnet 4.6 60.4% high Feb 2026 Official leaderboard
36 Kimi K3 60.4% max Jul 2026 Official leaderboard
37 Gemini 3.6 Flash 60.4% high Jul 2026 Official leaderboard
38 GPT-5.6 Luna 59.6% max Jul 2026 Official leaderboard
39 GPT-6 Luna 59.3% max Sep 2026 Official leaderboard
40 Muse Spark Open 57.5% xhigh Oct 2026 Official leaderboard
41 GPT-5.2 Pro 54.2% high Dec 2025 Official leaderboard
42 Gemini 3 Pro 54% Poetiq refinement Nov 2025 Official leaderboard
43 Grok 4.5 52.6% high Jul 2026 Official leaderboard
44 Gemini 3 Deep Think (Preview) 45.1% Feb 2026 Official leaderboard
45 Qwen 3.8 27B 42.4% xhigh Aug 2026 Official leaderboard
46 Inkling-Small 40.1% xhigh Jul 2026 Official leaderboard
47 Claude Opus 4.5 37.6% thinking, 64k Nov 2025 Official leaderboard
48 Inkling 36.5% Jul 2026 Official leaderboard
49 Gemini 3 Flash 33.6% high Dec 2025 Official leaderboard
50 Grok 4 29.4% J. Berman refinement Jul 2025 Official leaderboard
51 Beam 1.0 23.5% xhigh Oct 2026 Official leaderboard
52 GLM-5.2 22.8% Jun 2026 Official leaderboard
53 GPT-5.4 Mini 18.9% xhigh Mar 2026 Official leaderboard
54 GPT-5 Pro 18.3% Oct 2025 Official leaderboard
55 GPT-5.1 17.6% thinking, high Nov 2025 Official leaderboard
56 Claude Sonnet 4.5 13.6% thinking 32k Sep 2025 Official leaderboard
57 Kimi K2.5 11.8% Jan 2026 Official leaderboard
58 Gemini 3.5 Flash-Lite 10.3% high Jul 2026 Official leaderboard
59 GPT-5 9.86% high Aug 2025 Official leaderboard
60 Claude Opus 4 8.61% thinking 16k May 2025 Official leaderboard
61 o3 6.53% high Apr 2025 Official leaderboard
62 tiny-recursion-model 6.25% Oct 2025 Official leaderboard
63 o4-mini 6.11% high Apr 2025 Official leaderboard
64 Claude Sonnet 4 5.93% thinking 16k May 2025 Official leaderboard
65 GPT-5.4 Nano 5.69% xhigh Mar 2026 Official leaderboard
66 Grok 4 Fast 5.28% Sep 2025 Official leaderboard
67 Gemini 2.5 Pro Jun 2025 4.86% thinking 32k Jun 2025 Official leaderboard
68 o3-pro 4.86% high Jun 2025 Official leaderboard
69 GLM-5 4.86% Feb 2026 Official leaderboard
70 MiniMax-M2.5 4.86% Feb 2026 Official leaderboard
71 GPT-5 mini 4.44% high Aug 2025 Official leaderboard
72 Claude Haiku 4.5 4.03% thinking 32k Oct 2025 Official leaderboard
73 DeepSeek-V3.2 4.03% Dec 2025 Official leaderboard
74 o3-mini 2.99% high Jan 2025 Official leaderboard
75 GPT-5 nano 2.61% high Aug 2025 Official leaderboard
76 Gemini 2.5 Flash May 2025 2.54% thinking 24k May 2025 Official leaderboard
77 Nemotron 3.5 Super VL 1.94% medium Oct 2026 Official leaderboard
78 Gemini 2.0 Flash Dec 2024 1.30% Dec 2024 Official leaderboard
79 DeepSeek-R1 1.30% Jan 2025 Official leaderboard
80 codex-mini-2025-05-16 1.27% May 2025 Official leaderboard
81 Qwen3-235B-A22B-Instruct Jul 2025 1.25% Jul 2025 Official leaderboard
82 DeepSeek-R1 May 2025 1.12% May 2025 Official leaderboard
83 Claude 3.7 Sonnet 0.90% 8k Feb 2025 Official leaderboard
84 o1-mini 0.83% Sep 2024 Official leaderboard
85 Gemini 1.5 Pro Sep 2024 0.80% Sep 2024 Official leaderboard
86 GPT-4.5 0.80% Feb 2025 Official leaderboard
87 Grok-3 mini 0.42% low Apr 2025 Official leaderboard
88 GPT-4.1 0.42% Apr 2025 Official leaderboard
89 GPT-4o mini 0% Jul 2024 Official leaderboard
90 GPT-4o Nov 2024 0% Nov 2024 Official leaderboard
91 Llama 4 Maverick 0% Apr 2025 Official leaderboard
92 Llama 4 Scout 0% Apr 2025 Official leaderboard
93 Grok 3 0% Apr 2025 Official leaderboard
94 GPT-4.1 mini 0% Apr 2025 Official leaderboard
95 GPT-4.1 nano 0% Apr 2025 Official leaderboard
96 Magistral Medium 1.0 0% Jun 2025 Official leaderboard
97 Magistral Small 1.0 0% Jun 2025 Official leaderboard

Epoch AI has scores for 94 models. Labs report scores for 22 models in their cards. We plot every source in the charts below.

The best score, 95%, costs $1.12 a task

Runs on the cost frontier score higher than any run that's cheaper. GPT-6.1 Sol at high comes within 3.3 points of the best for $0.13 a task, 8 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

OpenAI DeepSeek Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, max GPT-6.1 Sol, max GPT-6.1 Sol, medium GPT-6.1 Sol, low DeepSeek V4 Flash 0731, max DeepSeek V4 Flash 0731, low GPT-6 Luna, medium 0% 20% 40% 60% 80% 100% $0.001 $0.01 $0.10 $1.00 $10 $100 Cost per task, log scale GPT-6 Astra, max GPT-6.1 Sol, medium
ModelEffortHarnessScoreCost per taskOutput tokens
GPT-6 Astra max 95% $1.12
GPT-6.1 Sol max 94.2% $0.25
GPT-6 Astra xhigh 93.3% $0.83
Claude Opus 5.5 high 93.3% $0.41
GPT-5.6 Sol max 92.5% $1.44
Claude Opus 5.5 xhigh 92.5% $0.67
GPT-6 Astra high 92.1% $0.67
GPT-6 Astra medium 92.1% $0.48
Claude Opus 5.5 max 91.7% $1.85
GPT-6.1 Sol xhigh 91.7% $0.18
GPT-6.1 Sol high 91.7% $0.13
Gemini 4 Argon high 91.3% $0.32
Claude Opus 5 max 90.4% $2.06
GPT-5.6 Sol xhigh 90% $1.04
Claude Fable 5.1 max 90% $4.49
Claude Fable 5.1 xhigh 90% $3.12
GPT-6 Sol max 89.6% $0.44
Claude Fable 5 max 89.2% $5.45
Gemini 3.8 Flash high 89.2% $0.40
Claude Fable 5.1 high 88.8% $1.67
Claude Fable 5 xhigh 88.3% $3.00
Claude Opus 5 high 88.3% $1.45
Claude Fable 5 high 87.5% $2.44
Claude Opus 5.5 medium 87.5% $0.34
Gemini 4 Argon medium 87.1% $0.14
GPT-6.1 Sol medium 86.7% $0.10
Claude Fable 5.1 medium 86.3% $1.22
GPT-5.6 Sol high 85.4% $0.74
GPT-6 Astra low 85.4% $0.42
GPT-5.5 xhigh 85% $1.87
Gemini 3 Deep Think 84.6% $14
GPT-5.5 Pro high 84.6% $11
Gemini 3.7 Flash high 84.6% $0.25
GPT-5.5 Pro xhigh 84.2% $11
GPT-5.6 Terra max 83.9% $1.09
Claude Sonnet 5.5 xhigh 83.9% $0.55
GPT-5.4 Pro xhigh 83.3% $16
GPT-5.5 high 83.3% $1.45
Gemini 3.8 Flash medium 82.9% $0.25
Claude Fable 5 medium 82.5% $1.93
Claude Fable 5.1 low 78.3% $0.95
GPT-6 Sol xhigh 78.1% $0.28
DeepSeek V4.1 Flash 1017 max 78.1% $0.16
Gemini 3.8 Flash low 77.5% $0.15
Gemini 3.1 Pro 77.1% $0.96
Claude Fable 5 low 76.8% $1.45
GPT-6.1 Sol low 76.7% $0.09
Claude Opus 4.7 max 75.8% $7.43
GPT-5.6 Terra xhigh 74.2% $0.69
GPT-5.4 xhigh 74% $1.52
DeepSeek V4.1 Flash 1017 low 73.6% $0.13
DeepSeek V4.1 Flash max 72.9% $0.13
GPT-5.2 Johan Land refinement 72.9% $39
Gemini 3.5 Flash high 72.1% $0.85
Claude Opus 4.8 high 72.1% $2.74
Claude Opus 4.8 medium 71.7% $2.39
GPT-5.5 medium 70.4% $0.86
Claude Opus 5.5 low 70.1% $0.24
Claude Opus 4.6 120k, high 69.2% $3.47
GPT-6 Sol high 68.9% $0.20
Claude Opus 4.6 120k, max 68.8% $3.64
DeepSeek V4.1 Flash 1017 high 68.6% $0.15
Claude Opus 4.7 high 68.3% $3.17
GPT-5.4 high 67.5% $1.02
Claude Opus 4.7 medium 67.5% $2.96
DeepSeek V4.1 Flash high 67.5% $0.13
Claude Sonnet 5.5 high 67.4% $0.28
GPT-5.6 Sol medium 67.1% $0.47
GPT-5.6 Terra high 67.1% $0.55
Grok 4.6 xhigh 67.1% $0.76
Claude Opus 4.6 120k, medium 66.3% $2.72
GLM-5.3-Flash max 65.8% $0.09
Grok 4.20 65.1% $0.92
Grok 4.6 high 65.1% $0.73
Claude Opus 4.6 120k, low 64.6% $2.25
Gemini 3.7 Flash medium 63.8% $0.12
Claude Opus 4.8 low 62.2% $1.68
Claude Opus 4.7 low 62.1% $2.38
DeepSeek V4 Flash 0731 max 61.4% $0.04
Grok 4.7 xhigh 61.4% $2.02
Grok 4.6 medium 61.3% $0.66
DeepSeek V4 Pro 0813 max 61.3% $0.60
Claude Haiku 5.5 xhigh 61.3% $0.08
DeepSeek V4.1 Flash low 60.6% $0.12
Claude Sonnet 4.6 high 60.4% $2.70
Kimi K3 max 60.4% $1.59
Gemini 3.6 Flash high 60.4% $0.61
DeepSeek V4 Pro 0813 high 59.7% $0.60
GPT-5.6 Luna max 59.6% $0.18
GPT-6 Astra none 59.6% $0.37
GPT-5.6 Luna max 59.5% $0.67
GPT-6 Luna max 59.3% $0.06
Grok 4.7 medium 58.8% $1.48
Claude Sonnet 4.6 max 58.3% $2.72
Grok 4.7 high 58.3% $1.55
GPT-6 Sol medium 57.8% $0.15
Muse Spark Open xhigh 57.5% $0.73
DeepSeek V4 Pro 0813 low 56.3% $0.37
DeepSeek V4 Flash 0731 high 56% $0.05
GPT-5.4 medium 55.4% $0.68
Kimi K3 high 55% $0.95
GPT-5.2 Pro high 54.2% $16
Gemini 3 Pro Poetiq refinement 54% $31
Gemini 3.7 Flash low 52.9% $0.08
GPT-5.2 xhigh 52.9% $1.90
Grok 4.5 high 52.6% $0.78
Grok 4.5 medium 52.6% $0.75
Gemini 3.6 Flash medium 50.4% $0.46
GLM-5.3-Flash high 50.1% $0.08
GPT-5.6 Luna xhigh 47.6% $0.54
Muse Spark Open high 47.5% $0.62
DeepSeek V4 Flash 0731 low 46% $0.02
GPT-5.6 Luna xhigh 43.5% $0.13
GPT-5.2 high 43.3% $1.39
GPT-5.6 Sol low 42.5% $0.32
Qwen 3.8 27B xhigh 42.4% $0.45
Gemini 4 Argon low 42.1% $0.07
GPT-6 Luna xhigh 41.9% $0.03
Inkling-Small xhigh 40.1% $0.23
GPT-5.2 Pro medium 38.5% $8.99
Claude Opus 4.5 thinking, 64k 37.6% $2.40
GPT-5.6 Terra medium 37.5% $0.27
Inkling 36.5% $0.64
Gemini 3 Flash high 33.6% $0.23
GPT-5.5 low 33.3% $0.35
GPT-5.6 Luna high 33.3% $0.06
Grok 4.5 low 33.1% $0.40
Inkling-Small high 33.1% $0.13
Claude Sonnet 5.5 medium 32.8% $0.16
GPT-6 Sol low 31.5% $0.10
GPT-6 Luna high 31.4% $0.02
Gemini 3 Pro 31.1% $0.81
Claude Haiku 5.5 high 30.7% $0.05
Gemini 3.6 Flash low 30.4% $0.22
Grok 4 J. Berman refinement 29.4% $30
GPT-5.6 Luna high 29.3% $0.29
GPT-5.4 low 29.2% $0.27
GLM-5.3-Flash low 27.9% $0.06
Grok 4.6 low 27.6% $0.21
Muse Spark Open medium 27.5% $0.25
GPT-5.2 medium 26.7% $0.76
Claude Sonnet 5.5 low 24% $0.11
Beam 1.0 xhigh 23.5% $0.61
Claude Opus 4.5 thinking, 16k 22.8% $0.79
GLM-5.2 22.8% $0.25
Qwen 3.8 27B low 22.8% $0.24
Beam 1.0 high 20% $0.35
GPT-5.4 Mini xhigh 18.9% $0.75
GPT-5.6 Terra low 18.8% $0.17
GPT-5 Pro 18.3% $7.14
GPT-6 Luna medium 18.1% $0.01
GPT-5.1 thinking, high 17.6% $1.17
Grok 4.7 low 16.7% $0.17
Grok 4 thinking 16% $2.17
Claude Opus 4.5 thinking, 8k 13.9% $0.48
Claude Sonnet 4.5 thinking 32k 13.6% $0.76
Inkling-Small medium 13.6% $0.06
GPT-5.4 Mini high 13.2% $0.56
Qwen 3.8 27B medium 13.2% $0.29
Gemini 3 Flash medium 12.8% $0.08
Kimi K3 low 12.4% $0.25
Muse Spark Open low 12.2% $0.11
Kimi K2.5 11.8% $0.28
Gemini 3.5 Flash-Lite high 10.3% $0.14
GPT-5 high 9.86% $0.73
GPT-5.2 low 9.72% $0.26
Claude Haiku 5.5 medium 9.44% $0.01
Gemini 3.5 Flash minimal 8.89% $0.11
Claude Opus 4 thinking 16k 8.61% $1.93
Claude Opus 4.5 thinking, none 7.78% $0.22
GPT-5 medium 7.49% $0.45
GPT-5.6 Luna medium 7.36% $0.02
GPT-5.6 Luna medium 7.36% $0.11
Claude Sonnet 4.5 thinking 8k 6.94% $0.23
Claude Sonnet 4.5 thinking 16k 6.94% $0.35
o3 high 6.53% $0.83
GPT-5.1 thinking, medium 6.53% $0.42
Beam 1.0 medium 6.39% $0.18
tiny-recursion-model 6.25% $2.10
o4-mini high 6.11% $0.86
Claude Sonnet 4 thinking 16k 5.93% $0.49
Claude Sonnet 4.5 thinking 1k 5.83% $0.14
GPT-5.4 Nano xhigh 5.69% $0.16
Grok 4 Fast 5.28% $0.06
Gemini 3.5 Flash-Lite medium 5.28% $0.04
GPT-5.6 Luna low 5.14% $0.07
Inkling-Small low 4.86% $0.06
Gemini 2.5 Pro Jun 2025 thinking 32k 4.86% $0.76
o3-pro high 4.86% $7.55
GLM-5 4.86% $0.27
MiniMax-M2.5 4.86% $0.17
GPT-6 Luna low 4.58% $0.005
Claude Opus 4 thinking 8k 4.52% $1.16
GPT-5 mini high 4.44% $0.20
GPT-5.4 Mini medium 4.44% $0.33
Gemini 2.5 Pro Jun 2025 thinking 16k 4.03% $0.71
GPT-5 mini medium 4.03% $0.06
Claude Haiku 4.5 thinking 32k 4.03% $0.38
DeepSeek-V3.2 4.03% $0.12
Inkling-Small minimal 4.03% $0.05
Claude Sonnet 4.5 3.75% $0.13
GPT-5.6 Luna low 3.61% $0.01
GPT-5.4 Nano high 3.61% $0.13
Gemini 3 Flash minimal 3.33% $0.02
o3-mini high 2.99% $0.55
o3 medium 2.98% $0.48
Gemini 2.5 Pro Jun 2025 thinking 8k 2.92% $0.44
Claude Haiku 4.5 thinking 16k 2.78% $0.14
Beam 1.0 low 2.78% $0.10
Claude Haiku 5.5 low 2.78% $0.008
Gemini 3.6 Flash minimal 2.64% $0.07
GPT-5 nano high 2.61% $0.03
Gemini 2.5 Flash May 2025 thinking 24k 2.54% $0.32
o4-mini medium 2.36% $0.23
Gemini 2.5 Flash May 2025 thinking 1k 2.16% $0.03
Gemini 2.5 Flash May 2025 thinking 8k 2.12% $0.20
Claude Sonnet 4 thinking 8k 2.12% $0.27
DeepSeek V4 Flash 0731 none 2.08% $0.004
DeepSeek V4.1 Flash 1017 none 2.08% $0.01
o3-mini medium 2.08% $0.28
o3-pro low 2.05% $2.23
o3 low 1.99% $0.23
Gemini 2.5 Flash May 2025 thinking 16k 1.98% $0.32
Nemotron 3.5 Super VL medium 1.94% $0.02
o3-pro medium 1.94% $4.74
GPT-5 low 1.94% $0.19
GPT-5.1 thinking, low 1.94% $0.13
GPT-5.4 Nano medium 1.94% $0.06
Gemini 2.5 Flash May 2025 1.69% $0.06
o4-mini low 1.67% $0.05
GPT-5 mini minimal 1.67% $0.009
Claude Haiku 4.5 thinking 8k 1.67% $0.09
GPT-6 Sol none 1.67% $0.07
GPT-5.4 Nano low 1.53% $0.01
Gemini 3.5 Flash-Lite low 1.53% $0.02
Qwen 3.8 27B none 1.53% $0.06
Nemotron 3.5 Super VL high 1.53% $0.04
Gemini 2.0 Flash Dec 2024 1.30% $0.004
DeepSeek-R1 1.30% $0.08
codex-mini-2025-05-16 1.27% $0.23
Claude Opus 4 1.27% $0.64
Claude Sonnet 4 1.27% $0.13
Qwen3-235B-A22B-Instruct Jul 2025 1.25% $0.004
Claude Haiku 4.5 1.25% $0.04
Claude Haiku 4.5 thinking 1k 1.25% $0.05
Gemini 3 Flash low 1.25% $0.02
DeepSeek V4.1 Flash none 1.25% $0.01
DeepSeek-R1 May 2025 1.12% $0.05
GPT-5.4 Mini low 1.11% $0.06
Claude 3.7 Sonnet 8k 0.90% $0.36
GPT-5 nano medium 0.88% $0.01
Claude Sonnet 4 thinking 1k 0.85% $0.14
GPT-5.6 Luna none 0.83% $0.008
Gemini 3.5 Flash-Lite minimal 0.83% $0.01
DeepSeek V4 Pro 0813 none 0.83% $0.04
Nemotron 3.5 Super VL none 0.83% $0.006
o1-mini 0.83% $0.19
GPT-5 mini low 0.83% $0.02
GPT-5.2 0.83% $0.08
Gemini 1.5 Pro Sep 2024 0.80% $0.04
GPT-4.5 0.80% $2.10
Claude 3.7 Sonnet 16k 0.70% $0.51
Grok-3 mini low 0.42% $0.01
GPT-4.1 0.42% $0.07
GPT-5.1 thinking, none 0.42% $0.06
Claude 3.7 Sonnet 1k 0.40% $0.14
GPT-4o mini 0% $0.01
GPT-4o Nov 2024 0% $0.08
o3-mini low 0% $0.06
Claude 3.7 Sonnet 0% $0.12
Llama 4 Maverick 0% $0.01
Llama 4 Scout 0% $0.006
Grok 3 0% $0.14
GPT-4.1 mini 0% $0.01
GPT-4.1 nano 0% $0.004
Claude Opus 4 thinking 1k 0% $0.75
Gemini 2.5 Pro Jun 2025 thinking 1k 0% $0.09
Magistral Medium 1.0 0% $0.11
Magistral Medium 1.0 thinking 0% $0.12
Magistral Small 1.0 0% $0.05
GPT-5 minimal 0% $0.06
GPT-5 nano low 0% $0.003
GPT-5 nano minimal 0% $0.003
Inkling-Small none 0% $0.01
GPT-6 Luna none 0% $0.004

11 models have held number one since its release in Mar 2025

Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard, Epoch AI’s runs and what labs claim in their cards.

OpenAI Anthropic xAI Google DeepMind
Own leaderboard Passed 85% in Apr 2026, by GPT-5.5
GPT-6 Astra 95%
Epoch AI runs Passed 85% in Apr 2026, by GPT-5.5
GPT-6 Astra 95%
Lab claims Passed 85% in Jul 2026, by Claude Opus 5
Claude Opus 5 90.4%

The best score passed 85% in Apr 2026

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Run or collected by Epoch AI Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-6 Astra Claude Fable 5 GPT-5.5 GPT-5.2 Gemini 3 Pro Grok 4 Claude Opus 4 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) GPT-6 Astra GPT-5.5

2 labs put it in a launch post in the last six months

Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.

xAI
ByteDance Seed
Google DeepMind
Anthropic
OpenAI
Ant Group
Meta
Thinking Machines
DateLaunch postScoreAs the post writes it
1 Sep 2026 Claude Fable 5.1 90 ARC-AGI-2
30 Jul 2026 Inkling-Small 40.1 ARC-AGI-2
24 Jul 2026 Claude Opus 5 90.4 ARC-AGI-2
19 May 2026 Gemini 3.5 Flash 72.1 ARC-AGI-2
14 May 2026 Ring-2.6-1T 66.2 ARC-AGI-V2 (xhigh)
16 Apr 2026 Claude Opus 4.7 75.8 ARC-AGI-2
8 Apr 2026 Muse Spark 42.5 ARC AGI 2 (Abstract Reasoning Puzzles (Public))
5 Mar 2026 GPT-5.4 Pro 83.3 ARC-AGI-2 (Verified)
19 Feb 2026 Gemini 3.1 Pro 77.1 ARC-AGI-2 (ARC Prize Verified)
17 Feb 2026 Claude Sonnet 4.6 58.3 Novel problem-solving ARC-AGI-2
14 Feb 2026 Seed2.0 Pro 37.5 ARC-AGI-2 (text) 37.5 / ArcAGI2-Image 43.3
12 Feb 2026 Gemini 3 Deep Think 84.6 ARC-AGI-2 (verified by the ARC Prize Foundation)
10 Feb 2026 Ring-2.5-1T 19.3 ARC-AGI-V2 (Pass@4)
5 Feb 2026 Claude Opus 4.6 68.8 Novel problem-solving ARC AGI 2
17 Dec 2025 Gemini 3 Flash 33.6 ARC-AGI-2 (ARC Prize Verified)
11 Dec 2025 GPT-5.2 Pro 54.2 ARC-AGI-2 (Verified), high
11 Dec 2025 GPT-5.2 52.9 ARC-AGI-2 (Verified) Abstract reasoning
24 Nov 2025 Claude Opus 4.5 37.6 Novel problem solving ARC-AGI-2 (Verified)
18 Nov 2025 Gemini 3 Pro 31.1 ARC-AGI-2 (ARC Prize Verified)
20 Aug 2025 seed-oss-36b-instruct 1.45 ArcAGI V2 (official evaluation set)
9 Jul 2025 Grok 4 15.9 ARC-AGI-2

12 system cards report it, from 5 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
1 Sep 2026 Claude Fable 5.1 system card 90 ARC-AGI-2 semi-private set, max effort, verified by ARC Prize Foundation (Table 8.1.A) Capabilities
30 Jul 2026 Inkling-Small model card 40.1 ARC-AGI-2 (5. Evaluations table) Capabilities
24 Jul 2026 Claude Opus 5 system card 90.4 ARC-AGI-2 semi-private set, max effort, verified by ARC Prize Foundation (90.42%) (Table 8.1.A, 8.14.1) Capabilities
23 Jun 2026 Seed2.1 Pro model card 62.5 ArcAGI2 (Table 11) Capabilities
13 Jun 2026 Ring-2.6-1T technical report 66.2 ARC-AGI-2 (Pass@2), xhigh (Table 6) Capabilities
19 May 2026 Gemini 3.5 Flash model card 72.1 ARC-AGI-2 (results table, p.4) Capabilities
16 Apr 2026 Claude Opus 4.7 system card 75.8 ARC-AGI-2, max thinking (Table 8.1.A, 8.11) Capabilities
19 Feb 2026 Gemini 3.1 Pro model card 77.1 ARC-AGI-2, ARC Prize Verified (results table, p.4) Capabilities
17 Feb 2026 Claude Sonnet 4.6 system card 58.3 ARC-AGI-2 (Verified), max effort (Table 2.1.A); 60.42 at high effort with 120k thinking tokens per ARC Prize (2.7) Capabilities
14 Feb 2026 Seed2.0 Pro model card 37.5 ARC-AGI-2, text input (Table 3). ArcAGI2-Image with text and image input 43.3 (Table 8) Capabilities
5 Feb 2026 Claude Opus 4.6 system card 68.8 ARC-AGI-2 (Verified), adaptive thinking, max effort (Table 2.3.A); ARC Prize reports 69.17 at high effort with 120k thinking tokens on the private set (2.9) Capabilities
24 Nov 2025 Claude Opus 4.5 system card 37.6 ARC-AGI-2 (Verified), semi-private set, 64k thinking (Table 2.3.A, 2.10) Capabilities

Versions

We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.

OpenAI ByteDance Seed xAI Anthropic Google DeepMind
ARC-AGI-3 Passed 85% in Sep 2026, by GPT-6 Astra
GPT-6 Astra 100%
ARC-AGI Passed 85% in Dec 2025, by GPT-5.2
Claude Fable 5 98.5%
ARC-AGI-2 Passed 85% in Apr 2026, by GPT-5.5
GPT-6 Astra 95%
ARC-AGI-3 Passed 85% in Sep 2026, by GPT-6 Astra
GPT-6 Astra 100%
Replaces ARC-AGI Replaced by ARC-AGI-3

More in Science and reasoning

All 89 in Science and reasoning
  • GPQA Diamond In use 198 graduate-level 'Google-proof' multiple-choice questions in biology, chemistry and physics. Gemini 4 Argon 96.4%
  • Terminal-Bench-Science In use 70 scientific research-workflow tasks across five science areas, run in Harbor containers. Hidden tests give the resolution rate. GPT-6 Astra 68.1%
  • FrontierScience In use Models solve physics, chemistry, and biology problems in an Olympiad track graded by short answer and a Research track graded by 10-point rubrics. Intern-S2-397B 87%
  • BioMysteryBench In use Bioinformatics problems on anonymized real datasets that require real analysis with tools. Rubric-graded, with human-solvable and human-difficult splits. Claude Opus 5 90.1%
  • HiPhO In use Models solve theory problems from 2024 and 2025 physics olympiad exams, some with diagrams. Official marking schemes grade answers and steps, mapped to medal levels. Gemini 3 Pro 78.3%
  • LABBench2 In use 11 biology research sub-tasks (literature, figures, tables, patents, trials and more) with tool access. Macro-averaged accuracy. GPT-5.2 60.9%