Same Model, 13.3% to 38.3%

OpenAI changed two API settings on ARC-AGI-3 and GPT-5.6 Sol went from 13.3% to 38.3% while spending six times fewer output tokens. Same model, same weights, same benchmark.