Enabling two API settings tripled GPT-5.6's ARC-AGI-3 scores

启用两项 API 设置使 GPT-5.6 在 ARC-AGI-3 基准测试得分提升三倍

GPT-5.6 Sol scored just 7.8% on ARC-AGI-3 because the official harness discarded private reasoning after each action and used rolling truncation that dropped older moves. Switching to retained reasoning and context compaction raised the public-set score from 13.3% to 38.3% while cutting output tokens by 6x. Human testers averaged about 48%. The post doesn't disclose full private-set results or whether the same settings help other models.

Why it matters: Official OpenAI post with concrete numbers and root-cause analysis, not marketing fluff. Capped below 85 because it's an engineering lesson rather than a capability breakthrough, and total score isn't disclosed. But 'the harness hurt the model' is directly useful for agent ben...