Calibration
AA-HLE → ARC-AGI-1
ARC-AGI-1 is estimated from AA-HLE with a linear curve fitted on 36 models measured on both: y = 0.7403·x + 0.5746, R² = 0.52, cross-validated error 8.1 pp. It is used for 2 estimates.
| Estimated model | AA-HLE | ARC-AGI-1 | Source |
|---|---|---|---|
| Phi-4 Multimodal Instruct | 5.0% | 61.2% | estimated ± 8.1 pp, low confidence |
| Llama 3.1 8B | 4.3% | 60.6% | estimated ± 8.1 pp, low confidence |
Fitted on
| Model | AA-HLE | ARC-AGI-1 |
|---|---|---|
| Claude Fable 5 | 55.5% | 98.5% |
| Claude Fable 5.1 | 59.1% | 97.5% |
| Claude Haiku 5.5 | 44.4% | 89.0% |
| Claude Opus 4.5 Thinking | 30.1% | 80.0% |
| Claude Opus 4.6 (Adaptive) | 39.9% | 93.0% |
| Claude Opus 4.8 | 48.7% | 92.5% |
| Claude Opus 5 | 54.9% | 97.5% |
| Claude Opus 5.5 | 61.4% | 97.5% |
| Claude Sonnet 4.6 | 13.3% | 86.0% |
| DeepSeek V4 Flash 0731 | 38.6% | 89.0% |
| DeepSeek V4 Pro | 41.0% | 90.0% |
| Gemini 3.5 Flash | 42.7% | 92.5% |
| Gemini 3.6 Flash | 40.8% | 91.2% |
| Gemini 3.7 Flash | 47.9% | 95.5% |
| Gemini 3.8 Flash | 47.8% | 98.5% |
| Gemini 3 Pro | 39.7% | 75.0% |
| Gemini 4 Argon | 57.1% | 98.5% |
| GPT-5.1 | 28.5% | 72.8% |
| GPT-5.2 | 37.7% | 86.2% |
| GPT-5.4 | 43.7% | 93.7% |
| GPT-5.4 mini | 28.1% | 63.7% |
| GPT-5.4 nano | 28.3% | 51.5% |
| GPT-5.5 | 45.8% | 95.0% |
| GPT-5.6 Luna | 39.5% | 88.0% |
| GPT-5.6 Sol | 49.5% | 96.5% |
| GPT-5.6 Terra | 42.9% | 96.5% |
| GPT-6.1 Sol | 52.9% | 96.5% |
| GPT-6 Astra | 54.7% | 97.5% |
| GPT-6 Luna | 38.5% | 86.7% |
| GPT-6 Sol | 47.9% | 95.5% |
| Grok 4.5 | 42.7% | 85.7% |
| Grok 4.6 | 42.9% | 87.0% |
| Grok 4.7 | 43.1% | 89.5% |
| Inkling 2 Small | 44.1% | 96.5% |
| Inkling-Small | 33.3% | 84.0% |
| Kimi K3 | 46.9% | 94.5% |