What the AI vendors are switching off, and when

Vendors call this deprecation. It means the thing you are using has a date on it. The vendor announces the date, sometimes months ahead, on a documentation page written for developers.

None of it is secret. It is simply on a page nobody in your business has been asked to read. So the announcement lands, nothing appears to happen, and the actual break arrives much later.

Below is every item we could tie to a vendor-published date and a primary source. Where a date sat behind a login, or came only from a news article, we left it out. An accurate short list is worth more than a long one you cannot check.

Two rows deserve a note. The fine-tuning wind-down is staged, so a business that fine-tuned a model last year is on a different date to one that never did. developers.openai.com ↗ And Claude Sonnet 4 matters more than its age suggests, because it is the last Claude model on Bedrock that routes across Asia rather than staying in Australia. If your data residency setting was copied from a tutorial, that is the model it was probably copied from. We cover that on which models can run in Australia.

You already track dates like these elsewhere. Your accounting software has a version that stops being supported. Your card terminal has a certificate that expires. Your insurance renews. Your AI vendor's roadmap belongs on the same list, and it moves faster than the rest.

The job is smaller than it sounds. One named person reads each vendor's deprecation page once a quarter, checks it against a list of what you actually run, and raises anything with a date inside twelve months. In a thirty-person business that is a reminder in a calendar, not a committee.

The failure mode is quiet, which is what makes it expensive. Nothing stops working on the day a product is announced as deprecated. Everything keeps running for another six months. Then the switch gets thrown on a Tuesday, the quoting tool starts returning errors, and the contractor who built it moved on in March.

If you paid somebody to build an AI workflow, this is a fair question to put to them in writing: which of these dates touch what you built for us, and what is the plan for each one. A good answer takes them ten minutes. No answer tells you something too.

The Claude dates above are the end-of-life dates printed on the Amazon Bedrock model cards. We did not verify the same dates against Anthropic's own API, and we do not assume they match. If you buy Claude directly rather than through Amazon, check with your provider which set of dates applies to you.

We did not find a single consolidated deprecation calendar published by Anthropic. Every Anthropic date on this page came from a separate per-model page on Amazon's documentation, one at a time.

The OpenAI fine-tuning date of 6 January 2027 covers creating new fine-tuning jobs. We could not find a published date for when inference on an existing fine-tuned model stops. If you run one, ask OpenAI in writing and keep the reply.

Neither vendor publishes a minimum notice period we could cite. So we cannot tell you how much warning the next one will come with, and neither can anybody else.

This page covers OpenAI and Anthropic only. We have not built the equivalent list for Microsoft Azure, Google or the smaller providers. When we do, it will appear here with its own dates.