Back to the blog

When your AI model gets an end date: a routine for the weeks before

· By Bee Flow · 5 min read

On 14 April 2026 Anthropic set 15 June as the retirement date for Claude Sonnet 4 and Claude Opus 4. What an IT lead can do in the weeks in between: a small test set, a pinned model ID, two dates in the calendar and one owner.

A desk calendar with one date circled in amber, next to a rack of plug-in modules where one module is being swapped for a new one, and a checklist card

On 14 April 2026 Anthropic notified developers that Claude Sonnet 4 and Claude Opus 4 will be retired on the Claude API on 15 June 2026. The model IDs concerned are claude-sonnet-4-20250514 and claude-opus-4-20250514, and the deprecations page names a recommended replacement for each. After the retirement date, the page says, requests to those models fail.

This is how the market works. Vendors release new models several times a year, retire older ones to make room, and say so in advance. Anthropic's page promises at least 60 days' notice for publicly released models, and this notice gives 62. Google announced on 18 February that gemini-2.0-flash and gemini-2.0-flash-lite will be shut down on 1 June 2026. Claude Haiku 3, announced in February, retires on 20 April. If your company uses AI through an API, a notice like this will reach you more than once a year, so it pays to have a routine for it.

Find out where the model is named

Before you can test anything, you need a list of every place a model ID is written down. In a company of fifty people it tends to be longer than expected: a support chatbot a supplier configured, a script that summarises tickets, an automation that sorts incoming mail, an internal tool with the ID in an environment variable, and the AI workspace staff use every day.

Anthropic's Console has a usage export broken down by API key and model, which is a quick way to see which keys still call a model that is going away. It does not show who owns each caller, so add that column yourself.

A test set of your own questions

The question to answer before 15 June is narrow: does the successor give answers you are happy with, on the work you use it for? A vendor's benchmark cannot tell you that.

A small test set can. Twenty to forty real tasks per use case is enough, each with the answer you consider good. Take them from real work: the customer mail that was hard to classify, the contract clause the model summarised well, the ticket where it asked a sensible follow-up question. Add a few where the current model got it wrong, so you notice when the new one does better. Strip personal data before an example goes into the set.

Keep the set next to the prompt it belongs to, together with the system prompt and settings. Then run it against the successor and compare. For classification and extraction a short script can count matches. For writing tasks, someone who knows the work reads both answers side by side without knowing which model wrote which.

Expect differences in length and tone even when the content is right. A prompt that leaned on the old model's habits may need one more sentence of instruction, and the test set is where you find that out, before your users do.

Anthropic's page gives the short version: test your applications with newer models well before the retirement date. With a test set ready, that takes an afternoon.

Pin the version, and know what an alias does

Many APIs accept two kinds of name. A pinned ID refers to one fixed model. An alias is a shorter name that the vendor may point at a different model later. On the Claude API, older models have aliases such as claude-sonnet-4-5 that resolve to the most recent dated snapshot, while from the 4.6 generation on the dateless ID is itself the pinned snapshot. Google's changelog shows what a move looks like: on 9 March 2026 gemini-3-pro-preview was shut down, and that name now points to gemini-3.1-pro-preview.

Our view: pin the ID in anything that runs unattended, such as automations, agents that act on incoming mail, and scheduled jobs. An alias spares you the switch but also skips your test, and you hear about the change from a user. For people chatting by hand an alias is fine, because they notice a difference straight away. Pinning does mean you meet every end date head on, which is why the next step matters.

Two dates in the calendar, and one name

The retirement date is the last day, not the plan. Put two earlier dates in a shared calendar. For a 15 June retirement:

  • 15 May: run the test set against the successor and write down the result.
  • 1 June: switch production over, with two weeks left to switch back if something unexpected turns up.

Then put a name next to those dates. Someone has to read vendor notices, keep the list of model IDs current, run the test set and make the call. In a small company that is usually the IT lead or whoever administers the AI tools, with the owner of each use case signing off on the result for their own work. If you have an AI policy, this belongs in it.

Anthropic's page says affected customers are notified by email, and other vendors work much the same way. Check that the address on each API account is a shared mailbox someone reads, and not the inbox of the colleague who opened the account two years ago.

A continuity plan for AI outages is a good home for all of this. The second model you keep ready for an outage is often the first one you test as a successor.

How Bee Flow handles a model change

In Bee Flow every agent has its own model setting, or follows the default model your organisation chooses. The prompt, the knowledge it draws on and its tools belong to the agent, so moving an agent to a successor is one setting and the rest stays as it was. Models from several providers can be connected side by side, so you can create a second agent with the same prompt on the successor and put your test questions to both before you switch. The platform page shows how agents and models fit together, and self-host or EU cloud shows where those models can run. If you would like to try it on one of your own use cases, get in touch.

#ai-models #ai-at-work