ARC-AGI-3 (Semi-Private)
General QA · 2026-03-25
ARC Prize's semi-private ARC-AGI-3 interactive reasoning evaluation, where agents must explore, infer goals, plan, and act in novel game-like environments. Scores are normalized against human action efficiency.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 99.9 |
| Gemini 4 Argon | 97.5 |
| GPT-6.1 Sol | 96.4 |
| Claude Sonnet 5.5 | 72.4 |
| Gemini 3.8 Flash | 35.0 |
| Claude Opus 5 | 30.2 |
| GPT-6 Sol | 23.0 |
| Grok 4.7 | 10.0 |
| GPT-5.6 Sol | 7.8 |
| DeepSeek V4.1 Flash | 3.3 |
| Claude Haiku 5.5 | 2.8 |
| Grok 4.6 | 2.1 |
| Opus 4.8 | 1.5 |
| GPT-5.6 Terra | 0.8 |
| GPT-6 Luna | 0.6 |
| Opus 4.6 | 0.5 |
| GPT-5.5 | 0.4 |
| Gemini 3.1 Pro Preview | 0.4 |
| Grok 4.5 | 0.3 |
| GPT-5.4 | 0.2 |
| Opus 4.7 | 0.2 |
| GPT-5.6 Luna | 0.2 |
| Grok 4.20 | 0.1 |