OpenAI's 28-Day Codex Streak: Caching, Edit History, and Locked Chats

OpenAI's 28-day Codex and Work shipping streak reached days 12 through 17 with a context caching upgrade on day 12 that reuses context across sub-agents and improves request routing, day 14 message editing in the Codex desktop app that discards all model outputs and tool calls after the edited point, and day 16 caching changes extending API cache to 12 hours and retaining cache indefinitely for logged-in Codex and Work users. The Decisions API, built on GPT-5.1 Codex, also became generally available for bounded-answer tasks such as probability estimates, classification and scoring, while PIN-protected locked chats roll out gradually. Days 13, 15 and 17 counted as resets, and OpenAI frames the streak as improving real usage quality rather than shipping a major feature daily.

OpenAI's 28-Day Codex Streak: Caching, Edit History, and Locked Chats OpenAI's 28-day Codex and Work shipping streak continues with caching upgrades, message editing, locked chats, and a wider Decisions API rollout. What is OpenAI’s 28-day Codex shipping streak? OpenAI committed to a 28-day run where every day either ships one concrete improvement to Codex or its Work product, or counts as an explicit “reset” day with no release. The stated goal is to improve real usage quality for people building with Codex and Work rather than chase big headline launches. Roughly a week into the second half of the streak, the pattern holds: most days bring small infrastructure or quality-of-life changes, punctuated by resets, with two features caching improvements and locked chats standing out as genuinely useful for daily users. TL;DR - OpenAI’s 28-day streak mixes small daily Codex and Work improvements with reset days when nothing ships, rather than promising a major feature every single day. - A context caching upgrade on day 12 reused context across sub-agents and improved request routing, which increases effective usage by hitting the cache more often and burning fewer tokens. - Day 16 extended API caching to 12 hours , and logged-in Codex and Work users now get cache retained indefinitely with no penalty for returning to an old conversation. - A message editing feature in the desktop app lets users edit a previous message in a thread, discarding all model outputs and tool calls that came after the edited point. - Locked chats are rolling out gradually, letting users PIN-protect conversations so they aren’t visible to anyone else who opens the account on a shared device. - The Decisions API , built on GPT-5.1 Codex “Luna” in the source material , became generally available and is positioned for bounded-answer tasks like probability estimates, classification, or scoring, rather than open-ended text generation. - Several of these updates compound: better caching plus persistent cache retention could meaningfully cut token costs for anyone running long or repeated Codex sessions. What shipped in the latest week of the streak? The most recent stretch covered days 12 through 17 with day 13, 15, and 17 counted as resets . Day 12 focused on effective usage gains through three mechanisms: reusing context across sub-agents, keeping cached context alive longer, and improving request routing so more requests actually hit the cache. None of these change what Codex can do, but they reduce the token overhead of using it, which in practice means more usable work out of the same plan or API budget. Day 14 introduced message editing in the Codex desktop app: users can go back and edit a message they sent earlier in a thread. Based on clarification from an OpenAI team member, editing a message discards all model outputs and tool calls that happened after that point in the conversation, rather than just redirecting the conversation going forward. That distinction matters for anyone relying on thread history for context, since it is closer to rewinding a conversation than branching it. Day 16 extended caching further: API requests can now cache for up to 12 hours, and logged-in Codex and Work users get cache retained indefinitely, with a guarantee of hitting the cache even after resuming a conversation hours later. This is arguably the most consequential change in the batch, since caching misses are a quiet but real cost driver for anyone running repeated or long-lived sessions against a model. How does the new caching system actually reduce costs? Caching in this context means the model doesn’t have to reprocess the same context tokens every time a session continues or a sub-agent reuses shared context. Before these changes, resuming an old conversation or coming back after a gap risked a cache miss, forcing the system to reprocess context from scratch at full cost. The day 12 update improved request routing specifically to increase the rate of cache hits, and the day 16 update removed the time penalty entirely for logged-in users, guaranteeing a cache hit even hours after the last message. For developers calling the API directly, the headline number is a 12-hour cache window, a meaningful jump for workloads that involve returning to the same context across a workday. For anyone using Codex or Work through a regular login rather than raw API calls, the retention is effectively permanent. The practical effect: less token re-processing, lower effective cost per session, and fewer surprises when picking a conversation back up later. What are locked chats and why do they matter? Locked chats let a user protect specific conversations behind a PIN, so anyone else who opens the account on a shared computer, for instance can’t read those chats without entering the code. The feature is rolling out gradually rather than to all accounts at once, which is standard for OpenAI’s staged releases. Seven tools to build an app. Or just Remy. Editor, preview, AI agents, deploy — all in one tab. Nothing to install. This fills an obvious gap for anyone using Codex or ChatGPT on a shared or semi-shared device: work accounts used by multiple people, family computers, or any situation where the chat history itself is sensitive strategy discussions, draft code with proprietary logic, personal information . It’s a low-complexity feature, but it addresses a real and common privacy gap that text-based AI tools have generally ignored. What is the Decisions API and who is it for? The Decisions API, built on GPT-5.1 Codex, is designed for a narrower class of problems than typical chat completions: questions with bounded, structured answers. Rather than returning open-ended text, it’s built to output things like an estimated probability, a choice from a fixed set of options, or a numerical score. OpenAI made it generally available in this update window, having previewed it earlier around its developer event and then expanded access in stages since. This positions it as a direct alternative to structured-output or classification-style tools elsewhere in the market, since the pitch is reliable, constrained outputs rather than free-form generation. For builders who need a model to pick from a defined set of categories, estimate likelihoods, or grade something on a scale, a dedicated API aimed at that use case avoids the extra engineering needed to force constrained behavior out of a general chat endpoint. Is this shipping pace actually meaningful for builders? Most of the individual updates in this batch are small. Message editing is a workflow convenience. A 12-hour API cache window and indefinite logged-in caching are genuinely useful but not flashy. Locked chats address privacy, not capability. Several days produced nothing at all. Taken item by item, none of this reads as a major leap. Taken together, though, the pattern is worth noting: OpenAI is optimizing the unglamorous parts of using Codex and Work day to day; cost efficiency through caching, basic data hygiene through locked chats, workflow friction through message editing, rather than racing to announce a new flagship model every week. For developers running Codex at any real volume, the caching changes alone could show up as a measurable drop in token spend. That’s a more durable kind of progress than a splashy demo, even if it makes for a quieter news cycle. Frequently Asked Questions What counts as a “reset” day in OpenAI’s 28-day streak? A reset day is a day when OpenAI doesn’t ship a Codex or Work improvement at all. The commitment was to ship a clear improvement or explicitly reset, rather than force out a minor change just to maintain a streak. Does the new caching system cost anything extra to use? Based on the update details, there’s no added penalty described for resuming an old conversation or hitting the cache. Logged-in Codex and Work users get the extended cache retention by default, and API users get up to a 12-hour cache window. How is the Decisions API different from a normal ChatGPT API call? A normal API call returns open-ended text. The Decisions API, built on GPT-5.1 Codex, is scoped to bounded-answer problems like picking from a fixed set of choices, estimating a probability, or producing a score, which suits classification or decision-support workloads better than free-text generation. Is the message editing feature the same as regenerating a response? No. Editing a previous message in a thread discards all model outputs and tool calls that occurred after that edited message, effectively rewinding the thread from that point rather than branching into a new path while keeping the old one. Are locked chats available to everyone right now? No. The feature is described as rolling out gradually, so availability depends on account and timing. Users who don’t see the option yet likely haven’t received it in their rollout group.