On 30 November, Anthropic switches off Claude Sonnet 4.5. On 11 December, OpenAI switches off GPT-5, GPT-5 mini and four other models in its API. On 14 January, GPT-4o and GPT-4.1 go the same way. If your business has a chatbot, an AI search box or an automated email triage built in the last two years, there is a fair chance it runs on one of them.
When that date comes, someone has to move your system to a newer model. The providers will tell you the newer model is better. It is also a different model, and it will answer your customers differently. Our verdict is to treat a model change as a change to your system, with a test before it goes live and a way back if it goes wrong.
The same name, different answers
The AI model is the one part of an AI system you do not own. You rent it from a provider, and its behaviour can shift under you.
Three researchers, Lingjiao Chen, Matei Zaharia and James Zou, measured this in 2023. They asked two versions of GPT-4, from March and June that year, the same questions. In March it identified prime numbers correctly 84% of the time. By June it managed 51% on the same questions. Their conclusion was that “the behavior of the ‘same’ LLM service can change substantially in a relatively short amount of time.” Other things moved too. GPT-4 followed instructions less well in June, while GPT-3.5 got better at the task GPT-4 got worse at.
The providers now address this directly. Anthropic says each model ID “identifies a pinned version of the model”, so a system that names an exact version gets the same model every time until that version is retired. Google’s advice is that “Most production apps should use a specific stable model.” Pinning protects you from surprise changes. It does not protect you from the retirement date.
When the update was the problem
The clearest case of a newer model making things worse came from OpenAI itself. On 24 and 25 April 2025, it rolled out an update to GPT-4o in ChatGPT. It began rolling the update back on 28 April. In OpenAI’s own words, the updated model had “skewed towards responses that were overly supportive but disingenuous.”
What matters for a business is why OpenAI’s testing missed it. Its follow-up post on 2 May said the update had added a new signal based on users’ thumbs-up and thumbs-down ratings. “We believe in aggregate,” it wrote, “these changes weakened the influence of our primary reward signal, which had been holding sycophancy in check.” Its offline evaluations “generally looked good”, and in A/B tests “the small number of users who tried the model liked it”. Some of its expert testers had said the model’s behaviour “felt” slightly off, but on the numbers OpenAI was watching, the update looked good. OpenAI’s conclusion was that its “offline evals weren’t broad or deep enough to catch sycophantic behavior.”
A UK example came 15 months earlier. In January 2024, a DPD customer got the company’s parcel chatbot to swear and to call DPD “the worst delivery firm in the world”. One post about it was viewed 800,000 times in 24 hours. The customer had asked the bot for exactly that, but it should have refused. DPD’s explanation, quoted by the BBC, was that “An error occurred after a system update yesterday. The AI element was immediately disabled and is currently being updated.” DPD said the AI had run successfully “for a number of years”. It went wrong after an update, and word spread after it was, in the BBC’s words, “spotted by a customer”.
Your business answers for what the bot says
A wrong answer from a chatbot is not the provider’s problem. It is yours.
In February 2024, a Canadian tribunal ruled on a case where Air Canada’s chatbot had told a customer that a bereavement fare could be claimed back after travel. That was wrong. Air Canada argued, in the tribunal’s summary, that “the chatbot is a separate legal entity that is responsible for its own actions.” The tribunal member called this “a remarkable submission” and found against the airline: “It makes no difference whether the information comes from a static page or a chatbot.” Air Canada was ordered to pay a total of C$812.02.
It is a small-claims decision from British Columbia, and it binds no UK court. The reasoning is the part worth keeping. A customer reading your chatbot sees your business, whichever model is producing the words that week.
The dates coming up
These are the main text-model retirements Anthropic, OpenAI and Google have announced for the next three months, read from their own deprecation pages on 24 October. OpenAI’s page also lists older image, voice and audio models.
| Model | Provider | Switched off | Notice given |
|---|---|---|---|
| Claude Sonnet 4.5 | Anthropic | 30 November 2026 | 30 September 2026 |
| GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3, o3-pro | OpenAI (API) | 11 December 2026 | 11 June 2026 |
| GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini | OpenAI (API) | 14 January 2027 | 14 October 2026 |
| Several preview voice and live models | Google (Gemini API) | 17 November 2026 | Preview models get at least two weeks |
Anthropic commits to “at least 60 days’ notice” for its public models, and its notice for Sonnet 4.5 was 61 days. OpenAI gave six months for GPT-5 and three for GPT-4o. Google says the dates in its table are “the earliest possible dates on which a model might be retired”.
The notices go to whoever holds the account with the provider. If that is the agency or freelancer who built your system, and they have since moved on, the email may be sitting in an inbox nobody reads.
It is not only the answers that can change. Anthropic’s API version 2026-10-21 rejects three settings developers commonly use to control how varied a model’s answers are (temperature, top_p and top_k), and returns an error instead. A system that sends them will stop working if it is moved to that version without its code being updated.
What a careful switch looks like
The providers’ own guidance agrees on the method. OpenAI recommends “Pinning your production applications to specific model snapshots” and “Building tests and evaluation suites that measure prompt behavior” for when you “change and upgrade model versions”. Anthropic tells developers to “Test your applications with newer models well before the retirement date.”
In practice, a careful switch has four parts:
A set of real questions with known right answers. Twenty to fifty is a reasonable start, taken from what your customers actually ask, including the awkward ones. We explain how these work in how to tell whether an AI system works.
A run of those questions against the new model before it goes live, with the answers compared against the old model’s answers rather than against a demo.
A named person who owns the date, holds the provider account, and reads its emails.
A way back. If the new model gets something wrong in the first week, someone can switch to the previous version while it is fixed, for as long as the provider still offers that version.
OpenAI’s experience shows why the questions matter more than the model’s reputation. The users in its A/B tests liked the April update. A team that only asks whether people like the new version is checking the wrong thing.
Questions to ask whoever built your AI system
1. Which model, and which exact version, does our system use? “The latest one” is not an answer. There should be a version name.
2. Is that version on a retirement list, and on what date? Every provider publishes one.
3. Who receives the provider’s retirement emails? If it is not you or someone you pay, fix that first.
4. What will you test before switching, and can we see the results? A useful answer names a set of questions and shows the old and new answers side by side.
What to do this month
Find out which model your system uses. If nobody can tell you, that is your first finding.
Check it against the three deprecation pages. Anthropic, OpenAI and Google each keep one, and the dates above are from them.
Collect twenty to fifty real questions your system gets, with the answers you would want. You will need them for this switch and for every one after it.
Book the switch with your supplier before December. A move tested in November is a routine job. One done in the last week before a shutdown is a scramble.
Frequently asked questions
Do AI models change after they are released?
A model can change under the same name. Stanford and Berkeley researchers found GPT-4 identified prime numbers correctly 84% of the time in March 2023 and 51% in June 2023. Providers now offer pinned versions that stay the same until they are retired, which is why your system should name an exact version.
When is GPT-4o being retired?
OpenAI retired GPT-4o from ChatGPT on 13 February 2026, as it announced on 29 January. In the API, which business systems use, GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini are scheduled to be switched off on 14 January 2027, according to OpenAI’s deprecations page.
When is Claude Sonnet 4.5 being retired?
Anthropic retires claude-sonnet-4-5-20250929 on 30 November 2026, with Claude Sonnet 5.5 as the recommended replacement. Anthropic notified developers on 30 September 2026.
Is a newer AI model always better for my business?
Not for every job. A newer model can answer your particular questions differently from the one you tested. OpenAI rolled back a GPT-4o update in April 2025 after it became, in its words, “overly supportive but disingenuous”, even though its own checks before release had looked good.
Is my business responsible for what its chatbot says?
Customers will treat it that way, and so did one tribunal. In Moffatt v Air Canada (February 2024), a Canadian tribunal rejected the airline’s argument that its chatbot was responsible for its own actions and ordered it to pay the customer. The decision is not binding in the UK.
Flux Dynamics is a fractional CTO who builds. If you are not sure which model your AI system runs on, or who would move it, tell us what you run and we will find out with you.