Best LLM for abstract reasoning (ARC-AGI-2)
- GPT-6 Astra leads
- 95.0%
- models ranked
- 72
- lead over #2
- +0.8 pts
ARC-AGI-2 · read October 21
As of October 21, 2026, OpenAI GPT-6 Astra leads 72 models on ARC-AGI-2 with 95.0%, ahead of OpenAI GPT-6.1 Sol (94.2%) and Anthropic Claude Opus 5.5 (93.3%).
The full best LLM for reasoning ranking combines this benchmark with others.
| Rank | Model | ARC-AGI-2 | Coding evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra OpenAI | 95.0% | 85.1 | $50.00 | |
| 2 | GPT-6.1 Sol OpenAI | 94.2% | 82.4 | $10.00 | |
| 3 | Claude Opus 5.5 Anthropic | 93.3% | 88.8 | $20.00 | |
| 4 | GPT-5.6 Sol OpenAI | 92.5% | 73.5 | $20.00 | |
| 5 | Gemini 4 Argon Google | 91.3% | 81.0 | $10.00 | |
| 6 | Claude Opus 5 Anthropic | 90.4% | 83.0 | $25.00 | |
| 7 | Claude Fable 5.1 Anthropic | 90.0% | 84.2 | $50.00 | |
| 8 | GPT-6 Sol OpenAI | 89.6% | 76.0 | $10.00 | |
| 9 | Claude Fable 5 Anthropic | 89.2% | 79.3 | $50.00 | |
| 10 | Gemini 3.8 Flash Google | 89.2% | 61.6 | $3.75 | |
| 11 | GPT-5.5 OpenAI | 85.0% | 71.3 | $30.00 | |
| 12 | Gemini 3.7 Flash Google | 84.6% | 58.0 | $3.75 | |
| 13 | GPT-5.5 Pro OpenAI | 84.6% | Not benchmarked | $180.00 | |
| 14 | GPT-5.6 Terra OpenAI | 83.9% | 63.9 | $12.00 | |
| 15 | GPT-5.4 Pro OpenAI | 83.3% | Not benchmarked | $180.00 | |
| 16 | Gemini 3.1 Pro Preview Google | 77.1% | 32.3 | $12.00 | |
| 17 | Claude Opus 4.7 Anthropic | 75.8% | 67.5 | $25.00 | |
| 18 | GPT-5.4 OpenAI | 74.0% | 64.4 | $15.00 | |
| 19 | DeepSeek V4.1 Flash DeepSeek | 72.9% | 62.5 | $0.78 | |
| 20 | Claude Opus 4.8 Anthropic | 72.1% | 65.0 | $25.00 |
How this list is built
Generated from the dataset on every refresh. Nobody picks the order by hand.
Models by their ARC-AGI-2 result, as the board publishes it, best first. The harder second set of ARC grid puzzles. Share of puzzles solved.
- OpenRouter, fetched
- Epoch AI benchmarking hub, fetched