OpenAI has introduced a configuration_update for GPT-6 Astra, enabling users to dynamically adjust the reasoning effort in the middle of a long conversation without rewriting the cached prompt prefix. This means that within the same conversation via the Responses API, users can now increase or decrease the reasoning effort at will, while the earlier context cache remains intact.
Previously, this was a hidden cost trap. The reasoning effort setting was embedded as a directive at the start of the context, so switching to a different effort level mid-conversation would invalidate a large portion of the cached data. A Codex user measured the impact: after switching the effort level, the cache hit rate fell from 98.5% to 65%, and only recovered to above 97% after the cache was rebuilt. This rebuild not only increased latency but also added computational overhead, especially in long conversations.
With the new update, users can temporarily set the reasoning effort to High or xhigh for difficult tasks, and then drop it back to Low for simpler ones, all while retaining the long context cache. For long-running agents, this is a practical improvement: it prevents the tens of thousands of tokens of cached context from being reprocessed as ordinary input just because the effort level changed. The change effectively eliminates the hidden cost of switching reasoning effort, making it more efficient and economical.