OpenAI added configuration_update to GPT-6 Astra on the Responses API. In the same long conversation, you can now raise or lower reasoning effort at any point without rewriting the cached prompt prefix. Hit a hard problem? Switch to High or xhigh. Back to an easy task? Drop to Low — and the long context from earlier stays reusable.
It used to be a sneaky money pit. Because reasoning effort sits at the front of the context as an instruction, switching to a different thinking tier mid-conversation could invalidate a large chunk of cache. One Codex user measured the cache hit rate dropping from 98.5% to 65% after a switch; it only climbed back above 97% after the next round rebuilt the cache.
That makes this genuinely useful for long-running agents. Hit something tricky and you can temporarily crank up the thinking dial; when the task is routine, dial it back down — without reprocessing hundreds of thousands of earlier tokens as ordinary input just because you changed gears.