OpenAI has announced ten dated deadlines between 23 October 2026 and 1 April 2027. Here is the calendar.

Everything below is taken from OpenAI's own deprecations page as read on 19 October 2026. Each entry was announced in public, with a date attached. None of it is a surprise, which is exactly what makes it dangerous: an announced deprecation only breaks the teams that were not looking.

OpenAI's page states its minimums. Generally available models get at least six months. Specialised variants of generally available models, such as chat variants, Codex variants and deep research variants, get at least three months. Preview models can be retired with much shorter notice, such as two weeks, and OpenAI says it does not recommend them for business critical production workloads unless you can migrate at short notice. It also reserves the right to move faster for safety or compliance reasons.

That is a useful filter when you read the calendar. A date six months out is a planning item. A date on a preview model is a deadline the day you read it.

On 7 September the list had seven dates ahead. Today it has ten, after the 24 September removal of the Videos API and Sora 2 passed, with no replacement named, and four dates were added: 31 October, 6 January, 14 January and 1 April. Behind them are the 1 October notices for text to speech, gpt-5.3-codex, gpt-5.1 and gpt-5.4-nano, and the 14 October notice for GPT-4o and the GPT-4.1 family.

One correction: the 7 September version named gpt-image-2 as the replacement for the 1 December image removals. OpenAI's page now names gpt-image-2.5-sunburst or gpt-image-2.5-flare, and this version follows it.

Rank them by where the call sits rather than by how soon the date is. A model string in your main request path gets caught in staging within a day. A model string in a nightly job, a webhook handler or an admin only feature can fail silently for weeks, because those paths fail into a queue or a log rather than into a user's face.

On that reading the October reasoning model shutdown is the sharpest of the set. o3-mini and o4-mini are reasoning models, which means teams typically route to them on an escalation path: the retry after a cheaper model failed validation, the "hard case" branch, the complex document handler. Those branches are the least exercised in testing and the most expensive when they stop working, because by definition they only run when something already needed care.

The 30 November platform shutdown is different in kind. Agent Builder, the Evals dashboard and reusable prompts are not code paths, they are places your process lives. Your evaluation history and your prompt library sit there. When that dashboard closes it takes your baselines with it, and you cannot rerun last quarter's evals to compare afterwards.

OpenAI documents an optional shutdown_date field on the List models endpoint, a YYYY-MM-DD string or null when none is announced, and it is also returned by Retrieve a model. One request shows every model your key can see that carries a date. It lists what is available to the key and not what your code calls, so pair it with a search.

Model identifiers are strings, which is what makes the search tractable. A single grep across your repository for the affected identifiers will find the large majority of call sites in a couple of minutes. Two places that grep will miss, and that are worth checking by hand: configuration that lives outside the repo, such as environment variables, feature flags, or a model name stored in a database row per customer; and anything that builds a model string by concatenation, which is rare but does happen in routing code that assembles a family name and a size.

Our own Cost Comparator listed o3-mini and o4-mini as current models with no caveat, and we publish a dedicated pricing page for o3-mini. Both have had a published shutdown date since 22 April 2026.

Worse, our weekly automated price check flagged both models on 5 September because their prices no longer appeared on OpenAI's current pricing page, and we recorded that as a likely false positive. The tool was right and the person reading it was wrong. The entries now carry the shutdown date, and the correction is published in the price change log rather than quietly applied.

We mention it because the mechanism that failed here is the same one that fails everywhere else: not the absence of a signal, but a human deciding the signal was noise.