BenchmarksReasoning
ARC-AGI-2
The ARC-AGI-2 abstract-reasoning puzzle benchmark (semi-private set), the harder static successor to ARC-AGI-1.
ReasoningaccuracyHigher is better
What is ARC-AGI-2?
The ARC-AGI-2 abstract-reasoning puzzle benchmark (semi-private set), the harder static successor to ARC-AGI-1. evals.report tracks reported ARC-AGI-2 scores with the model, source, status, date, and run caveats attached — official leaderboard scores, vendor-reported launches, and clearly labeled community runs.
Top reported ARC-AGI-2 score: GPT-5.5 — 85% (accuracy).
| Model | Lab | Score↓ | Source model | Status | Date | |
|---|---|---|---|---|---|---|
| GPT-5.5 | OpenAI | 85% | GPT-5.5 (xHigh) | Official | Apr 23, 2026 | Details |
| Gemini 3 Deep Think | Google DeepMind | 84.58% | Gemini 3 Deep Think (2/26) | Official | Dec 4, 2025 | Details |
| GPT-5.5 Pro | OpenAI | 84.58% | GPT-5.5 Pro (High) | Official | Apr 23, 2026 | Details |
| GPT-5.4 Pro | OpenAI | 83.33% | GPT-5.4 Pro (xHigh) | Official | Mar 5, 2026 | Details |
| Gemini 3.1 Pro Preview | Google DeepMind | 77.08% | Gemini 3.1 Pro (Preview) | Official | Feb 19, 2026 | Details |
| Claude Opus 4.7 | Anthropic | 75.83% | Claude 4.7 (Max) | Official | Apr 16, 2026 | Details |
| GPT-5.4 | OpenAI | 73.95% | GPT-5.4 (xHigh) | Official | Mar 5, 2026 | Details |
| Claude Opus 4.8 | Anthropic | 72.08% | Claude Opus 4.8 (High) | Official | May 28, 2026 | Details |
| Gemini 3.5 Flash | Google DeepMind | 72.08% | Gemini 3.5 Flash (High) | Official | May 19, 2026 | Details |
| Claude Opus 4.6 | Anthropic | 69.17% | Claude Opus 4.6 (120K, High) | Official | Feb 5, 2026 | Details |
| Grok 4.20 beta reasoning | xAI | 65.14% | Grok 4.20 (Reasoning) | Official | Mar 9, 2026 | Details |
| Claude Sonnet 4.6 | Anthropic | 60.42% | Claude Sonnet 4.6 (High) | Official | Feb 17, 2026 | Details |
| GPT-5.2 | OpenAI | 52.91% | GPT-5.2 (xHigh) | Official | Dec 11, 2025 | Details |
| Claude Opus 4.5 | Anthropic | 37.64% | Opus 4.5 (Thinking, 64K) | Official | Nov 24, 2025 | Details |
| Gemini 3 Flash | Google DeepMind | 33.61% | Gemini 3 Flash Preview (High) | Official | Dec 17, 2025 | Details |
| Gemini 3 Pro | Google DeepMind | 31.11% | Gemini 3 Pro | Official | Nov 18, 2025 | Details |
| GLM-5.2Open | Z.ai | 22.78% | GLM-5.2 | Official | Jun 16, 2026 | Details |
| GPT-5.1 | OpenAI | 17.64% | GPT-5.1 (Thinking, High) | Official | Nov 12, 2025 | Details |
| Grok 4 | xAI | 15.97% | Grok 4 (Thinking) | Official | Jul 9, 2025 | Details |
| Claude Sonnet 4.5 | Anthropic | 13.61% | Claude Sonnet 4.5 (Thinking 32K) | Official | Sep 29, 2025 | Details |
| Kimi K2.5Open | Moonshot AI | 11.81% | Kimi K2.5 | Official | Jan 27, 2026 | Details |
| GPT-5 | OpenAI | 9.86% | GPT-5 (High) | Official | Aug 7, 2025 | Details |
| o3 | OpenAI | 6.53% | o3 (High) | Official | Apr 16, 2025 | Details |
| o4-mini | OpenAI | 6.11% | o4-mini (High) | Official | Apr 16, 2025 | Details |
| Claude Sonnet 4 | Anthropic | 5.93% | Claude Sonnet 4 (Thinking 16K) | Official | May 22, 2025 | Details |
| GLM-5Open | Z.ai | 4.86% | GLM-5 | Official | Feb 11, 2026 | Details |
| GPT-5 mini | OpenAI | 4.44% | GPT-5 Mini (High) | Official | Aug 7, 2025 | Details |
| Claude Haiku 4.5 | Anthropic | 4.03% | Claude Haiku 4.5 (Thinking 32K) | Official | Oct 15, 2025 | Details |
| DeepSeek V3.2Open | DeepSeek | 4.03% | Deepseek V3.2 | Official | Dec 1, 2025 | Details |
Each row reports the model’s accuracy on ARC-AGI-2. Click a row for the full run context.