OpenAI 28-Day Product Sprint Day 16: Codex and Work cache permanently retained, API extended to 12 hours.

Beating AI News Flash: OpenAI's Head of Product and Platform, Thibault Sottiaux, announced Day 16 of the 28-day product sprint. This time, the focus is on solving the context caching issue for Agents running over extended periods. For logged-in Codex and ChatGPT Work users, OpenAI announced permanent retention of prompt caching. Even if work resumes after several hours or old sessions are reopened, cache hits are guaranteed, eliminating the loss of cache reuse benefits due to long pauses.

The API has also been upgraded in sync. OpenAI has extended the optional cache duration for the Responses API from the default 30 minutes to 12 hours. Developers can cache system instructions, tool definitions, and common contexts for the long term, allowing Agents to reuse previous computation results when executing tasks hours later. The two durations can also be mixed: fixed instructions are retained for 12 hours, while new conversation content is cached for only 30 minutes.

However, extending API caching comes at an additional cost. The initial write price for 12-hour caching is 2.5 times that of regular input, double the default 30-minute cache. On subsequent cache hits, input prices for most new models are reduced by 90%, while GPT-6.1 Sol is reduced by 95%. Therefore, long caching is more suitable for Agents that repeatedly call the same context, rather than short tasks executed only once.