Switching reasoning effort mid-session can affect more than just the answer.
In one measured Codex session, a new effort level dropped the cache hit rate from above 92% to about 66% on the next turn.
That meant 4,096 fewer input tokens were served from the cache.
