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

OpenAI reported that enabling two API settings tripled scores for GPT-5.6 on the ARC-AGI-3 benchmark. The gains came from retaining reasoning and enabling compaction, which together raised measured performance while also improving efficiency on that evaluation.

On the product side, the change is configuration, not a new model release. Retaining reasoning keeps intermediate reasoning available across the run instead of discarding it. Enabling compaction compresses or condenses that retained state so longer multi-step work stays within usable context. On ARC-AGI-3, which stresses abstract, multi-step problem solving, those two levers changed how the same GPT-5.6 endpoint uses tokens and carries state through a task.

For engineers building agentic or multi-step systems, this is a concrete lever: benchmark and product behavior can shift hard from API flags alone. If you evaluate models without matching reasoning retention and compaction, you may understate what GPT-5.6 can do on ARC-AGI-3-style workloads and mis-size latency, cost, and failure modes in production agents.

In market terms, the write-up frames progress on ARC-AGI-3 as partly an inference- and session-configuration problem, not only a training problem. Competitors and evaluators who publish scores without documenting equivalent settings risk non-comparable leaderboards; builders comparing vendors need the same two controls in the test harness if they want a fair read against OpenAI’s reported 3x result.

Practical takeaway: when you next run GPT-5.6 on hard multi-step or abstraction tasks, explicitly test reasoning retention on and off and compaction on and off, log score, token use, and wall time, and pin the winning pair in your client config. Watch for whether OpenAI documents defaults, limits, and cost side effects of those two settings, and whether later evals require them to reproduce the ARC-AGI-3 triple.