ARC-AGI leaderboard
- Claude Fable 5 leads
- 98.5%
- models measured
- 73
- at the top
- Tied
- median
- 86.5%
Leader on ARC-AGI · read October 22
On October 22, 2026, 78 of our catalog's models had a ARC-AGI result, and Anthropic Claude Fable 5 led with 98.5%.
Grid puzzles where the model has to infer a rule from a few examples and apply it. Share of puzzles solved.
Role in the index
ARC-AGI is shown as published and does not feed any of our indexes. Read the methodology
| # | Model | Result | Cost per task | Evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5Anthropic | 98.5% | Not published | 79.3 | $50.00 | |
| 2 | Claude Opus 5.5Anthropic | 98.5% | Not published | 88.8 | $20.00 | |
| 3 | Gemini 3.8 FlashGoogle | 98.5% | Not published | 61.6 | $3.75 | |
| 4 | Gemini 4 ArgonGoogle | 98.5% | Not published | 81.0 | $10.00 | |
| 5 | GPT-6 AstraOpenAI | 98.5% | Not published | 85.1 | $50.00 | |
| 6 | GPT-6.1 SolOpenAI | 98.5% | Not published | 82.4 | $10.00 | |
| 7 | Gemini 3.1 Pro PreviewGoogle | 98.0% | $0.5223 | 32.3 | $12.00 | |
| 8 | Claude Fable 5.1Anthropic | 97.5% | Not published | 84.2 | $50.00 | |
| 9 | Claude Opus 5Anthropic | 97.5% | Not published | 83.0 | $25.00 | |
| 10 | GPT-5.6 SolOpenAI | 97.5% | $0.40 | 73.5 | $20.00 | |
| 11 | GPT-5.5 ProOpenAI | 96.5% | $4.53 | Not benchmarked | $180.00 | |
| 12 | GPT-5.6 TerraOpenAI | 96.5% | $0.55 | 63.9 | $12.00 | |
| 13 | Gemini 3.7 FlashGoogle | 95.5% | Not published | 58.0 | $3.75 | |
| 14 | GPT-6 SolOpenAI | 95.5% | Not published | 76.0 | $10.00 | |
| 15 | GPT-5.5OpenAI | 95.0% | $0.73 | 71.3 | $30.00 | |
| 16 | DeepSeek V4.1 FlashDeepSeek | 94.5% | Not published | 62.5 | $0.84 | |
| 17 | Kimi K3Moonshot AI | 94.5% | $0.7705 | 62.3 | $15.00 | |
| 18 | GPT-5.4 ProOpenAI | 94.5% | $7.21 | Not benchmarked | $180.00 | |
| 19 | Claude Opus 4.6Anthropic | 94.0% | Not published | 57.2 | $25.00 | |
| 20 | GPT-5.4OpenAI | 93.7% | $0.62 | 64.4 | $15.00 | |
| 21 | Claude Opus 4.7Anthropic | 93.5% | $1.41 | 67.5 | $25.00 | |
| 22 | Claude Opus 4.8Anthropic | 92.5% | $2.332 | 65.0 | $25.00 | |
| 23 | Gemini 3.5 FlashGoogle | 92.5% | $0.4278 | 39.8 | $9.00 | |
| 24 | Gemini 3.6 FlashGoogle | 91.2% | Not published | 62.5 | $3.75 | |
| 25 | GLM 5.3 FlashZ.ai | 91.0% | Not published | 57.7 | $0.50 | |
| 26 | DeepSeek V4 Pro 0813DeepSeek | 90.5% | Not published | 56.1 | $1.98 | |
| 27 | GPT-5.2 ProOpenAI | 90.5% | $11.6542 | Not benchmarked | $168.00 | |
| 28 | Grok 4.7xAI | 90.2% | Not published | 66.7 | $6.00 | |
| 29 | Grok 4.20xAI | 89.5% | $0.35 | 33.3 | $2.50 | |
| 30 | Claude Haiku 5.5Anthropic | 89.0% | Not published | 70.7 | $0.50 | |
| 31 | DeepSeek V4 Flash 0731DeepSeek | 89.0% | Not published | 51.8 | $0.396 | |
| 32 | GPT-5.6 LunaOpenAI | 88.0% | $0.32 | 58.0 | $1.20 | |
| 33 | Qwen3.8 27BQwen | 87.5% | Not published | 74.2 | $2.55 | |
| 34 | Grok 4.6xAI | 87.5% | Not published | 65.7 | $6.00 | |
| 35 | Grok 4.5xAI | 87.2% | $0.31 | 58.6 | $6.00 | |
| 36 | GPT-6 LunaOpenAI | 86.7% | Not published | 57.3 | $0.50 | |
| 37 | Claude Sonnet 4.6Anthropic | 86.5% | $1.1908 | 57.3 | $15.00 | |
| 38 | GPT-5.2OpenAI | 86.2% | $0.96 | 55.9 | $14.00 | |
| 39 | Gemini 3 Flash PreviewGoogle | 84.7% | $0.1743 | 43.7 | $3.00 | |
| 40 | T Inkling SmallThinking Machines | 84.0% | Not published | 27.3 | $1.20 | |
| 41 | Claude Opus 4.5Anthropic | 80.0% | $1.47 | 44.0 | $25.00 | |
| 42 | T InklingThinking Machines | 79.5% | $0.3005 | 31.5 | $4.05 | |
| 43 | GLM 5.2Z.ai | 77.0% | $0.1916 | 61.1 | $4.40 | |
| 44 | GPT-5.1OpenAI | 72.8% | Not published | 40.4 | $10.00 | |
| 45 | GPT-5 ProOpenAI | 70.2% | Not published | Not benchmarked | $120.00 | |
| 46 | GPT-5OpenAI | 65.7% | $0.5087 | 38.8 | $10.00 | |
| 47 | Kimi K2.5Moonshot AI | 65.3% | $0.15 | 38.6 | $2.85 | |
| 48 | Claude Sonnet 4.5Anthropic | 63.7% | $0.5193 | 39.5 | $15.00 | |
| 49 | MiniMax M2.5MiniMax | 63.7% | $0.07 | 37.2 | $1.20 | |
| 50 | GPT-5.4 MiniOpenAI | 63.7% | $0.51 | 35.8 | $4.50 | |
| 51 | o3OpenAI | 60.8% | $0.5002 | Not benchmarked | $8.00 | |
| 52 | o3 ProOpenAI | 59.3% | $4.16 | Not benchmarked | $80.00 | |
| 53 | o4 MiniOpenAI | 58.7% | $0.4058 | Not benchmarked | $4.40 | |
| 54 | DeepSeek V3.2DeepSeek | 57.0% | $0.08 | 30.0 | $0.88 | |
| 55 | GPT-5 MiniOpenAI | 54.3% | $0.116 | 34.2 | $2.00 | |
| 56 | Gemini 3.5 Flash LiteGoogle | 53.5% | Not published | 49.0 | $2.50 | |
| 57 | GPT-5.4 NanoOpenAI | 51.5% | $0.13 | 44.8 | $1.25 | |
| 58 | Claude Haiku 4.5Anthropic | 47.7% | $0.2558 | 35.6 | $5.00 | |
| 59 | GLM 5Z.ai | 44.7% | $0.17 | 41.7 | $3.20 | |
| 60 | Gemini 2.5 ProGoogle | 41.0% | $0.4839 | 19.1 | $10.00 | |
| 61 | Claude Opus 4RetiredAnthropic | 35.7% | $1.2496 | Not benchmarked | $75.00 | |
| 62 | o3 MiniOpenAI | 34.5% | $0.3989 | Not benchmarked | $4.40 | |
| 63 | Gemini 2.5 FlashGoogle | 33.3% | $0.2134 | Not benchmarked | $2.50 | |
| 64 | o1OpenAI | 30.7% | Not published | Not benchmarked | $60.00 | |
| 65 | Claude 3.7 SonnetRetiredAnthropic | 28.6% | $0.33 | Not benchmarked | $15.00 | |
| 66 | o1-proOpenAI | 23.3% | Not published | Not benchmarked | $600.00 | |
| 67 | R1 0528DeepSeek | 21.2% | $0.0464 | Not benchmarked | $2.18 | |
| 68 | GPT-5 NanoOpenAI | 20.7% | $0.0124 | Not benchmarked | $0.40 | |
| 69 | R1RetiredDeepSeek | 15.8% | $0.06 | Not benchmarked | $2.50 | |
| 70 | o1-miniRetiredOpenAI | 14.0% | Not published | Not benchmarked | $4.40 | |
| 71 | Qwen3 235B A22B Instruct 2507Qwen | 11.0% | $0.0025 | Not benchmarked | $0.75 | |
| 72 | GPT-4.1OpenAI | 5.5% | $0.039 | Not benchmarked | $8.00 | |
| 73 | Grok 3RetiredxAI | 5.5% | $0.0931 | Not benchmarked | $15.00 | |
| 74 | GPT-4oOpenAI | 4.5% | $0.05 | Not benchmarked | $10.00 | |
| 75 | Llama 4 MaverickMeta Llama | 4.4% | $0.0078 | Not benchmarked | $1.00 | |
| 76 | GPT-4.1 MiniOpenAI | 3.5% | $0.0078 | Not benchmarked | $1.60 | |
| 77 | Llama 4 ScoutMeta Llama | 0.5% | $0.0041 | Not benchmarked | $0.50 | |
| 78 | GPT-4.1 NanoOpenAI | 0.0% | $0.0021 | Not benchmarked | $0.40 |
Only models in our catalog are listed, and each appears once with its best published run. 215 rows were read from the source on October 22, 2026, and rows that match no catalog model are left out. Cost per task is the spend on one puzzle, in US dollars, as Epoch AI records it beside the result.