6 min read
The AI model behind your favorite automation will be switched off one day, and you will probably get the warning by email at a moment you are not paying attention. That is not a hypothetical. According to Anthropic’s model deprecations page, Claude Sonnet 4.5 retires from the Claude API on November 30, 2026, and any request still pointed at it after that date will fail. Meanwhile, the newest model on the same page, Claude Fable 5.2, went live on November 9.
New models arrive every few weeks; old ones quietly expire. For a one-person business, the right response is not to chase every release. It is to write a short model-switching policy once, so the next retirement notice is a 30-minute chore instead of a broken Monday. Here is how to think about it and what to put in yours.
First, know whether this affects you at all
Model retirements mostly matter in one place: anything that calls an AI model directly with a specific model name. In plain terms, a “model” is the specific AI engine doing the thinking, and each version has its own name, like a product SKU. Chat apps such as ChatGPT, Claude or Gemini generally move you to newer models on their own, so a retirement there is usually invisible.
You are affected if any of these are true:
- You built an automation in Zapier, Make or n8n with an AI step where you picked a model from a dropdown
- A freelancer built you a small script, chatbot or website feature that uses an API key (a password-like code that lets software talk to the AI company directly)
- You use a third-party app that lets you choose “which AI model” in its settings
If none of those apply, you can stop reading here and enjoy your coffee. If even one applies, keep going.
The three facts that should shape your policy
Retirements come with notice, but not a lot
Anthropic says it gives at least 60 days’ notice before retiring publicly released models, and its table lists a “not sooner than” date for every active model. Fable 5.2, for example, will not retire sooner than November 9, 2027. OpenAI keeps a similar public list on its deprecations page. Sixty days sounds generous until the email lands in December, during your busiest season.
“Deprecated” and “retired” mean different things
A deprecated model still works but is no longer recommended, and it has a retirement date. A retired model no longer works at all. Anthropic’s own documentation warns that deprecated models are likely to be less reliable than active ones. In practice, deprecation is your signal to move, not retirement.
A newer model is not automatically a drop-in replacement
Newer models are usually smarter, but they can be wordier, format answers differently or follow instructions more literally. An automation that relied on a model always replying with a single word (“yes” or “no”) can break when a newer model replies with a full sentence. The newer engine is better; your setup just was not built for it. Our explainer on context windows covers another way models differ in how much text they can handle at once.
Your one-page model-switching policy
Write this down once, in a note you will actually find again. Five lines are enough.
- Inventory. List every place you use an AI model by name: which tool, which model, what it does, and what breaks if it stops. Most solo owners have between one and five entries.
- Watch list. Bookmark the deprecation pages for the AI companies you use, and check them once a quarter. Put the date on your calendar.
- Switch trigger. Move to a newer model when your current one is marked deprecated, not on the day it retires. Do not move just because something new launched.
- Test set. Keep five real examples of inputs your automation handles (five customer emails, five product descriptions, five invoices) along with the output you consider correct. Run them through the new model before switching.
- Rollback. Note how to switch back, and keep the old setting visible until the new one has run cleanly for a week (as long as the old model is still available).
That is the whole policy. It turns a vague worry into a checklist.
How to run the 30-minute switch
When a deprecation email arrives, here is the routine:
- Find the recommended replacement. Deprecation pages usually name one. Anthropic’s table, for example, recommends Claude Sonnet 5.5 as the replacement for Sonnet 4.5.
- Duplicate the automation rather than editing the live one. In Zapier, Make and n8n, copying a workflow takes a click.
- Change only the model in the copy. Leave the prompt exactly as it is for the first test.
- Run your five test examples and compare against your saved correct outputs. Look for format changes first, because that is what most often breaks the next step.
- Adjust the prompt if needed. If the new model is wordier, add a line like “Reply with only one word: YES or NO.” Then retest.
- Switch over, and watch it for a week. Then delete the old copy.
If your automation costs money per run, check the new model’s price before switching. A smarter model that costs three times as much per step can quietly change your monthly bill; our breakdown of every AI tool you pay for is a good place to record the change.
Why not just always use the newest model?
It is tempting. Launches are exciting, and each one promises better results. But for a business, predictable beats impressive. An automation that sends invoices or answers customer questions should behave the same way on Tuesday as it did on Monday. Chasing every release means retesting constantly, and the gains on simple business tasks are often small.
The better habit is to try new models in a chat window first, on your real work, and switch your automations only when there is a clear reason: deprecation, a meaningful price drop, or a capability you actually need. We made a similar argument when Google AI Plus lost its Pro models: change follows need, not headlines.
The bigger picture: AI vendors are becoming suppliers
The useful mental shift is this: an AI model is a supplier, not a utility. Like any supplier, it can change terms, raise prices or discontinue a product line. Anthropic has published commitments on model deprecation and preservation, which is a reasonable sign of how seriously the industry now treats this. But no commitment replaces your own plan.
Solo owners already know how to manage suppliers. You keep a backup, you watch for discontinued items, and you test a new product before you put it in front of a customer. Apply the same instincts here and model retirements stop being emergencies.
Do this today
Open a note called “AI models I depend on.” Write down every tool where you chose a model by name. If Claude Sonnet 4.5 is on that list, you have until November 30. If it is not, you have just finished the hardest part of your policy. How many entries are on your list?
