ARC-AGI-1 Semi-Private
General QA
ARC-AGI-1 results on the 100-task Semi-Private evaluation set, including ARC Prize verified evaluations. The official scorer gives each task the fraction of its test pairs solved by either permitted attempt and averages those fractional task scores; it is not a count of binary task successes. This held-out cohort is distinct from the 400-task public evaluation set and has its own capability factor.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5 | 98.5 |
| Claude Opus 5.5 | 98.5 |
| Gemini 3.8 Flash | 98.5 |
| Gemini 4 Argon | 98.5 |
| GPT-6.1 Sol | 98.5 |
| GPT-6 Astra | 98.5 |
| Gemini 3.1 Pro Preview | 98.0 |
| Claude Fable 5.1 | 97.5 |
| Claude Opus 5 | 97.5 |
| GPT-5.6 Sol | 97.5 |
| Claude Sonnet 5.5 | 96.5 |
| GPT-5.5 Pro | 96.5 |
| GPT-5.6 Terra | 96.5 |
| Gemini 3 Deep Think | 96.0 |
| Gemini 3.7 Flash | 95.5 |
| GPT-6 Sol | 95.5 |
| GPT-5.5 | 95.0 |
| DeepSeek V4.1 Flash 1017 | 94.5 |
| DeepSeek V4.1 Flash | 94.5 |
| GPT-5.4 Pro | 94.5 |
| Kimi K3 | 94.5 |
| Opus 4.6 | 94.0 |
| Opus 4.7 | 93.5 |
| GPT-5.4 | 92.7 |
| Opus 4.8 | 92.5 |
| Gemini 3.5 Flash | 92.5 |
| Gemini 3.6 Flash | 91.2 |
| GLM-5.3 Flash | 91.0 |
| DeepSeek V4 Pro 0813 | 90.5 |
| GPT-5.2 Pro | 90.5 |
| Grok 4.7 | 90.2 |
| Grok 4.20 | 89.5 |
| Claude Haiku 5.5 | 89.0 |
| DeepSeek-V4-Flash-0731 | 89.0 |
| Muse Spark Open | 89.0 |
| GPT-5.6 Luna | 88.0 |
| Grok 4.6 | 87.5 |
| Qwen3.8 27B | 87.5 |
| Grok 4.5 | 87.2 |
| GPT-6 Luna | 86.7 |