According to @OpenAI, enabling two API settings tripled ARC-AGI-3 scores with 6x fewer tokens after fixing a harness that blocked learning recall.

According to emollick, GPT-5.6 Sol became SoTA on ARC-AGI-3 via two harness settings enabling multi-window reasoning and compaction, per OpenAI analysis.

According to @emollick, Anthropic’s Claude Opus 5 hits 30.2% on ARC-AGI-3, surpassing the prior 7.8% SOTA and showing novel behaviors, per ARC Prize.