When a model retires

On 23 October OpenAI switched off o1 and o1-pro in its API, and with them a row of older models: GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, o3-mini, o4-mini and the first snapshot of GPT-4o. Nine days earlier it had told developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will follow on 14 January. None of this is unusual. Old models retire, and the new ones are usually cheaper and more capable. But for every company that has tuned a product or a workflow to one model, it is a reminder that the model was rented, and that it has an end date.

What went, and what comes next

OpenAI announced the October shutdown on 22 April, six months ahead, saying it wanted “to improve reliability and make it easier for developers to choose the right models”. Fine-tuned versions went too, among them those built on GPT-4 and GPT-3.5 Turbo. For most of the retired models the deprecations page names a substitute from the newer GPT-5.6 family.

The January list is shorter but reaches further. GPT-4o arrived in May 2024 as OpenAI’s flagship, and GPT-4.1 was a family made for developers and the API. OpenAI recommends one replacement for all four, gpt-6-luna, which it launched in September. On OpenAI’s own terms this is routine. “As we launch safer and more capable models, we regularly retire older models,” the page says, and software built on them “may need occasional updates to keep working.”

Three months, not six

What stands out is the notice. OpenAI’s policy promises at least six months for generally available models, at least three for specialised variants such as chat and Codex versions, and as little as two weeks for models marked preview. Two entries posted on 1 October state their notice plainly, one of them “with six months’ notice”. The entry of 14 October gives about three months and no reason.

GPT-4o and GPT-4.1 carry no preview label and are not chat or Codex variants, so by our reading they belong in the six-month class. The New Stack reached the same conclusion and noted that “there’s no explanation for why GPT-4o and GPT-4.1 are getting only half that time.”

To be fair to OpenAI, its policy allows faster timelines for safety or compliance reasons, and three months is not short by the standards of the industry. Anthropic promises “at least 60 days’ notice” for publicly released models, and gave Claude Sonnet 4.5 two months, from 30 September to 30 November. Google’s Gemini page calls its shutdown dates the “earliest possible dates on which a model might be retired” and names no minimum at all. The lesson is about all of them: the notice period is the provider’s to set, and it can change.

What breaks when a model goes

Changing the model name in the code takes a minute. Everything around it takes longer. The New Stack reported that gpt-6-luna handles reasoning and tool calls differently from GPT-4.1: in the Chat Completions interface it calls tools only with reasoning switched off, and with reasoning on, familiar settings such as temperature are no longer available. Its verdict: “swapping the model ID may be the easiest part of the migration.”

The deeper changes do not show up as errors. Prompts refined over months to one model’s habits now meet a model with other habits. Answers come back longer or shorter, in another format or another tone, and a check that passed for a year can quietly start to fail.

Cost moves as well, mostly in the customer’s favour. OpenAI lists gpt-6-luna at 0.10 dollars per million input tokens, against 2.50 dollars for GPT-4o, although the rate rises for prompts longer than 272'000 tokens. Fine-tuned models are the hardest case. New self-serve fine-tuning jobs end on 6 January, and a model trained on a company’s own data retires with the base it was built on.

How to plan for it

The first defence is an evaluation set of your own: real cases from your own work, with answers you know are right, run against every candidate before it goes live. Keep it, and the prompts, in your own repository. The same deprecations page shows why. OpenAI is also retiring its Evals platform, whose existing evals turn read-only on 31 October, and its reusable prompt objects, telling developers to move that content into their application code.

The second is to separate the product from the model. If the model is chosen in one place, behind one interface, a switch becomes a configuration change and a test run, not a rebuild. Pinning a dated snapshot keeps behaviour from drifting, but it does not stop the date from arriving.

The third is simply to watch. OpenAI’s List models interface reports a shutdown date for every model that has one, so it can go into monitoring like the expiry of a certificate, with a named person who answers for it.

The fourth is for the work that must stay stable for years: a model whose weights you hold. Nobody can switch it off for you, and you decide when to upgrade. That has its own costs, from hardware to the updates you run yourself, and the question of where an open model comes from, which we looked at earlier this month. Anthropic has committed to preserving the weights of its models over the long term, and hopes “to make past models publicly available again”. That is welcome, but it does not keep a product running today.

Where we stand

Werk, our app platform, is live, and the apps we build on it with clients use the best model for each job, or a private model with Eigna. Eigna, a product of Eigna Labs that we set up for clients as its partner, is in early access. It is an open model adapted to a client’s domain, running in its own cloud or on its own servers. A new version goes live only when it tests better than the last one, and it can be switched back in one step. The model changes when the company decides, not when a provider does.