The class of a model is legible from its ID string, which is what makes this policy something a lint rule can check instead of something a person has to remember. Anything with `preview` in the name sits on the short clock, which OpenAI puts at possibly as little as two weeks [3]. A `-codex` suffix, a `-chat-latest` suffix, or a deep research variant is a specialized variant of a generally available model and gets at least three months [2]. A plain GA snapshot gets at least six [1]. The migration budget attached to a pin is a property of the string in your config.
The log is where the floor gets tested. On August 26, 2026, OpenAI notified users of `whisper-1` and the `gpt-4o-transcribe` family that those models leave the API on February 26, 2027 [5]. Legacy audio, realtime and transcription families were noticed on July 20, 2026, for January 20, 2027 [6]. Older GPT-5 and o3 snapshots were noticed on June 11, 2026, for December 11, 2026 [7]. Day of month is preserved in all three, so the shutdown date is notice plus six calendar months, computed rather than negotiated [9]. "At least six months" is being implemented as six months. Nothing in the dated entries shows a longer window, so plan to the floor.
Six months is roughly 26 weeks, which puts a two-week preview retirement at about a thirteenth of the GA runway [10]. A variant pin gets half [18]. The page also says outright not to use preview models for business-critical production workloads unless you can migrate on short notice [4], which is the unusual case of documentation settling an argument before you have it.
For six months to be your planning number, the retirement has to be an ordinary one. Where safety or compliance forces a faster timeline, the published minimums do not apply and the commitment is as much notice as reasonably possible [8]. That clause carries no number, so the six months is a floor for routine retirement and unbounded below for the other kind. Two states also matter for how you read your own dashboards: a model becomes deprecated the instant it is announced, and the shutdown date is a separate thing that ends access [11]. Nothing in your config changes when that happens.
The notice table classifies models. It does not classify platform surfaces, and on June 3, 2026, three of those went at once: Agent Builder, the Evals platform, and reusable prompt objects [13][14][15]. No minimum notice period is published for any of them [19]. The migration path for prompt objects is to move the content into application code [15]. Agent Builder users are pointed at the Agents SDK or ChatGPT Workspace Agents, with ChatKit still available [13]. Evals users are pointed at Promptfoo, which is somebody else's tool [14].
If a model you depend on shuts down anyway, the documented escape hatch is provisioning dedicated capacity for continued access, arranged through the sales team [16]. That price comes out of a sales conversation, with no published rate card to check it against. So the honest way to size a migration budget is to read the model strings in your config and count the hyphens now, well ahead of any announcement email [17].
