ARC-AGI-3 实测：GPT-6.1 Sol 高推理更省步数

ARC Prize @arcprize原推发布于 2026.10.01

在 X 上打开原推

Like GPT-6 Astra, GPT-6.1 Sol made better decisions at higher reasoning levels on ARC-AGI-3, requiring fewer actions to complete levels, which reduced the total inference cost. For example, on sk48 using the provider adapter (which preserves opaque reasoning state between turns and enables auto compaction), GPT-6.1 Sol with max reasoning worked out how to reposition the colored blocks around an obstacle sooner, while low spent much longer revisiting blocked moves and testing controls that didn't help. That helped it complete the game in 463 actions versus 1,078 with low reasoning. Watch the replay: https://arcprize.org/replay/c299de73-6d1c-45f5-be74-c3552c90af73