Two API parameters tripled OpenAI's scores on ARC-AGI-3. Enabling reasoning item retention and conversation compaction in the GPT-5.6 API allowed the model to carry forward its chain-of-thought across multi-turn problem attempts, instead of rebuilding context from scratch each time.
The gains were not just in accuracy. Compaction cut token overhead significantly, meaning the efficiency improvement compounds: better scores at lower cost per run. ARC-AGI-3 is a demanding benchmark specifically designed to resist pattern memorization, so performance lifts here carry more signal than on saturated datasets.
The full post details the exact API calls, the benchmark structure, and the before-and-after score breakdown. If you are running any long-horizon reasoning evaluations or agentic loops on OpenAI models, the implementation specifics are worth reading directly.
[READ ORIGINAL →]