How enabling two settings tripled our scores on the ARC-AGI-3 benchmark
Source: OpenAI Blog29/07/2026, 12:00
Two API settings dramatically improved GPT-5.6 performance on the ARC-AGI-3 benchmark, tripling the scores. The settings preserve reasoning and enable compaction, boosting overall efficiency. This discovery demonstrates how simple optimizations can have exponential impact on AI model performance.