OpenAI Codex Head Tibo: API Supports 12-Hour Caching, Logged-In Users Retain Cache Permanently
Tibo, the head of OpenAI Codex and ChatGPT, announced that the API now supports a 12-hour prompt caching option, while logged-in users can retain cache permanently in Codex and Work.
Users returning to old conversations or resuming after several hours will no longer face penalties and are guaranteed cache hits; previously, OpenAI developer accounts announced a 12-hour caching option for the API, suitable for persistent agents to reuse context within a day.
This update is part of Tibo's 28-day challenge, marking the 16th day of improvements. On the 12th day, context reuse across sub-agents, extended cache duration, and optimized routing were implemented to increase cache hit rates and enhance effective usage.
The permanent cache for logged-in users allows shared instructions, tools, and context to be used for longer periods at a lower cache input rate, even with request intervals of several hours; the API side offers a 12-hour option for developers to choose from.
Tibo has previously announced Codex usage resets and feature improvements multiple times via X, including disk storage optimization and reduced model capacity errors; this cache extension directly addresses the issue of context loss after long interruptions.
The official OpenAI developer account also confirmed the 12-hour cache, emphasizing that persistent agents can reuse context throughout the day at a lower rate; the indefinite cache for logged-in users further reduces the cost of conversation recovery.
From a market mechanism perspective, the cache extension is driven by product efficiency, potentially increasing user frequency of long conversations and agent tasks; beneficiaries are developers and Codex users relying on long-term context, while competitors with shorter caching strategies may face pressure.
Source: Public Information
ABAB AI Insight
Tibo, as the head of OpenAI Codex and ChatGPT, has long publicly announced usage resets and feature iterations through X, forming a 28-day challenge format; he has previously optimized context reuse and cache hits multiple times, and this time extends the logged-in user cache from a limited period to permanent, continuing his path of enhancing user experience through public updates.
The capital path is to reduce the computational cost of repeated inputs by extending cache time, allowing persistent agents and long-term projects to operate at lower rates; the motivation is to increase effective usage and user stickiness, reducing context reconstruction due to interruptions, making the platform more competitive in agent and long-term task scenarios.
This is similar to the historical continuous optimization of conversation state and caching strategies, or cloud services adjusting persistent storage and computing resources; currently, we are in a stage of popularizing AI agents and long-term interactions, where cache persistence becomes a key improvement to reduce friction.
Essentially, it represents a technological substitution and cost structure change: reducing repeated computation and context reconstruction through permanent caching, with the mechanism converting short-term caches into long-term reuse, lowering the costs of workflows that rely on continuous context, making users more willing to maintain long conversations, while alternatives with shorter cache windows face efficiency disadvantages.
ABAB News · Cognitive Laws
The longer the cache time, the lower the reconstruction cost after interruptions.
Permanently caching turns temporary conversations into continuous workflows.
Guaranteeing hits reduces penalties more than simply extending time.