GPT-5.6 Persisted Reasoning: Reuse Thinking Across Turns

When OpenAI pushed GPT-5.6 to general availability on July 9, 2026, most of the noise went to the three-tier model family (Sol, Terra, Luna) and to programmatic tool calling. A quieter addition matters just as much if you run multi-turn agents: persisted reasoning, controlled by a single new field, reasoning.context.

Here is the problem it solves. Reasoning models spend tokens thinking before they answer. In a normal multi-turn session those reasoning tokens are thrown away between turns, so on turn two the model re-derives the same plan, the same constraints, and the same assumptions it already worked out on turn one. You pay for that thinking twice. On a ten-turn planning session with a stable goal, you pay for it ten times.