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

The blog post highlights two configurable settings in the GPT-5.6 API: one that keeps the model's chain-of-thought reasoning intact for final generation, and another that compacts that reasoning trace to reduce token usage. Enabling both settings together tripled performance on ARC-AGI-3, a benchmark designed to test abstract reasoning and generalization. The reasoning retention prevents loss of intermediate logic, while compaction lowers output length and cost, making high-level reasoning tasks more practical for production use. This suggests developers can achieve major gains on reasoning-heavy workloads without custom fine-tuning, simply by adjusting inference parameters.