AI Product Operations

How Codex prompt caching behaves when you change settings

Six matched tests of what Codex reuses after a model or reasoning change, updated for GPT-6, where a reasoning change no longer has to cost the cached conversation.

Jorge Alcantara/August 20, 2026/7 min read
A tree-shaped model of Codex prompt caching shows a shared cached base, two conversation branches, a colder switch turn, and the possible return of an earlier branch.

More field notes

Keep the useful details in the loop.

Subscribe for tested notes on AI systems, product decisions, and the context that makes both work better.

All posts