What Happens When the Model Behind Your Automation Is Retired?

What Happens When the Model Behind Your Automation Is Retired?

As of September 2026, it fails. Not slowly, and not with a warning in your logs. Anthropic's documentation is blunt about it: requests to retired models will fail. Every AI provider now runs a published retirement schedule, and an automation pinned to a model past its date simply stops returning answers.

This is the quiet operational risk in almost every AI workflow we see. Teams spend weeks tuning a prompt, wire it into a lead router or a content pipeline, and then never look at the model name again. Model names have expiry dates now, and nobody on the marketing team knows that.

We think this deserves more attention than it gets, because the fix is cheap and the failure is expensive. Here is how the schedules actually work and what we do about them.

What Is the Difference Between a Deprecated Model and a Retired One?

Deprecated means still working but no longer recommended. Retired means gone. Anthropic uses four states on its model deprecations page: Active, Legacy, Deprecated, and Retired. A deprecated model still answers requests and carries a published retirement date. A retired one returns an error.

OpenAI draws the same line with different words. Its deprecations documentation says that when a model or endpoint is announced as being deprecated, it immediately becomes deprecated, and uses "sunset" or "shutdown" for the moment it becomes inaccessible.

The practical takeaway is that deprecation is your warning window, not your problem. The problem is the shutdown date, and it is always on a calendar somewhere you are not looking.

How Much Notice Do the Big Providers Actually Give?

Less than most teams assume, and it varies by tier. OpenAI commits to at least 6 months of notice before retiring generally available models, at least 3 months for specialised variants such as chat, codex and deep research models, and roughly 2 weeks for preview models. Anthropic commits to at least 60 days of notice before retiring publicly released models.

Sixty days is not long if the automation belongs to a marketing team with a full quarter already planned. Six months is comfortable, but only if somebody reads the email. In our experience the notification lands in a developer inbox and the person who owns the workflow never sees it.

OpenAI is explicit about preview models: it does not recommend using them for business critical production workloads. That sentence is worth reading twice, because we regularly find preview model names sitting in production automations that a team built during an excited week of experimentation.

Are These Schedules Real or Theoretical?

They are real and they run on time. Anthropic retired Claude Sonnet 3.7 on 19 February 2026, having announced the deprecation on 28 October 2025. It retired Claude Opus 4.1 on 5 August 2026, announced on 5 June 2026, which is a 61 day window. Google shut down the Gemini 1.5 family on 29 September 2025 and the Gemini 2.0 Flash family on 1 June 2026.

OpenAI has October 23, 2026 on the calendar for both gpt-3.5-turbo-0125 and gpt-4-0613, and 26 August 2026 for the Assistants API. Whisper-1 and gpt-4o-transcribe are listed for 26 February 2027.

Look at that list and notice how many of those model names appeared in tutorials and starter templates that teams copied. A workflow built from a 2024 blog post is very likely pointing at something already switched off.

Should You Pin a Model Version or Follow the Latest Alias?

Pin it. An alias that silently moves to a newer model means your prompt output changes one morning with no deploy and no ticket. That is the harder failure to diagnose, because nothing in your system changed.

A pinned version fails loudly on a known date instead. Loud and scheduled beats silent and random every time. It turns a mystery into a calendar entry.

The cost of pinning is that you own the upgrade. That is the right trade for anything touching customer data, published content, or money. For a throwaway internal script, an alias is fine.

What Actually Breaks When You Swap Models?

Rarely the API call. Almost always the output shape and the tone. A new model may be better at reasoning and worse at obeying a format instruction you wrote for the old one. Structured output schemas hold up well. Loose instructions like "reply with three short lines" do not.

Cost and latency shift too, sometimes in your favour. A newer small model can beat an older large one on both. We have seen teams delay a forced migration for months and then discover the replacement was cheaper all along.

Prompts that leaned on a quirk of one model are the real casualty. If your prompt works because of how a specific version handled a specific edge case, it was always fragile. Retirement just reveals it, which is one reason we treat prompt versioning and testing as core infrastructure rather than a nice to have.

How Do You Find Every Automation That Uses an Old Model?

Start with billing, not with code. Anthropic's Console has a Usage page with an Export button that produces a CSV broken down by API key and model. That single file usually tells a team more about its own AI footprint than an afternoon of asking around.

Then grep your repositories for model name strings, and check the places code does not live: Zapier steps, Make scenarios, n8n nodes, Airtable automations, and whatever a clever operations person built one Friday. Those are where orphaned model names hide.

Write the result down in one place. Every automation, the model it calls, the person who owns it, and the retirement date. We keep this alongside our other AI automation documentation, because an undocumented automation is an outage waiting for a date.

Can Parameters and APIs Be Deprecated Too?

Yes, and this catches people out. Anthropic has deprecated the temperature, top_p and top_k parameters on Claude Opus 4.7 and later, where setting them to a non-default value returns a 400 error. Its Python SDK from version 1.0 removes them outright, so passing them raises a TypeError.

Whole endpoints go too. OpenAI's Assistants API is scheduled for 26 August 2026. An automation built on a deprecated endpoint needs rearchitecting, not a one line model swap, and that is a very different amount of work to plan for.

So the audit question is not only "which model does this call". It is also "which API surface and which parameters does this depend on".

How Do We Build Automations That Survive This?

Three habits, and none of them are clever. We keep the model name in configuration rather than scattered through code, so a swap is one edit. We write a small set of example inputs with expected outputs for every automation, so a model change can be checked in minutes rather than argued about. And we make the automation fail loudly, with an alert to a human, rather than returning something empty.

The example set matters most. Ten real inputs with agreed good answers turn "does the new model still work" from a feeling into a test. That is the same discipline we describe in evaluating an AI automation before production.

We also put the retirement dates in the same calendar the team already uses. A date in a vendor changelog is not a reminder. A date in your shared calendar with an owner attached is.

What Will Model Retirement Look Like a Year From Now?

Expect the cadence to stay fast and the notice windows to stay roughly where they are. Providers retire models to free capacity for new ones, and Anthropic says so directly in its documentation. There is no sign that older models become permanently available, though Anthropic has committed to preserving model weights long term.

So the durable answer is not to chase stability from vendors. It is to build automations where changing the model is a small, tested, boring event. Teams that do this upgrade in an afternoon. Teams that do not find out when a form stops routing leads.

If you have AI automations running in your marketing stack and nobody can tell you which models they call, that audit is a good week's work and we are happy to walk through how we would run it. You can find us at phoenix.studio.