benchgap
Calibration

GPQA → ARC-AGI-3

ARC-AGI-3 is estimated from GPQA with a offset logistic curve fitted on 12 models measured on both: y = 0.0000 + (0.7656 − 0.0000) / (1 + exp(−200.00·(x − 0.9548))), R² = 0.99, cross-validated error 9.3 pp. It is used for 4 estimates.

Estimated modelGPQAARC-AGI-3Source
Qwen3.5 397B88.4%0.0%estimated ± 9.3 pp, low confidence
Qwen3.6-27B87.8%0.0%estimated ± 9.3 pp, low confidence
Qwen3.5-122B-A10B86.6%0.0%estimated ± 9.3 pp, low confidence
Qwen3.6-35B-A3B86.0%0.0%estimated ± 9.3 pp, low confidence

Fitted on

ModelGPQAARC-AGI-3
Claude Opus 4.691.3%0.5%
Claude Opus 4.794.2%0.2%
Claude Opus 4.893.6%1.5%
GPT-5.492.8%0.2%
GPT-5.593.6%0.4%
GPT-5.6 Luna92.3%0.2%
GPT-5.6 Sol94.6%7.8%
GPT-5.6 Terra92.9%0.8%
GPT-6.2 Astra97.0%70.3%
GPT-6.2 Sol96.0%54.9%
GPT-6 Astra96.0%62.7%
Inkling 2 Small93.0%0.3%