According to CoinMeta, OpenAI has added a configuration_update for GPT-6 and Astra, allowing for the adjustment of reasoning and effort at any time within the same long conversation without overwriting the previously cached prompt prefix. Users can temporarily increase their thinking intensity when dealing with difficult problems and reduce it for simpler tasks, avoiding the invalidation of previous caches due to such adjustments. Real user tests have shown that after switching gears, the cache hit rate dropped from 98.5% to 65%, but it recovered to over 97% only after a new round of caching was established. This change is very practical for agent that runs for extended periods.