How enabling two settings tripled our scores on the ARC-AGI-3 benchmark

OpenAI reported that enabling two API settings on GPT-5.6 dramatically improved its performance on the ARC-AGI-3 benchmark. The settings are reasoning retention and compaction. By keeping reasoning steps and compacting the output, the model achieved scores three times higher than before, while also improving efficiency. The discovery suggests that simple configuration changes can unlock significant performance gains without modifying the model itself.