Here's something most owners never hear when they buy an AI tool: the engine underneath has an expiration date.
The AI behind your chatbot, your phone assistant or your automation is a "model," and the companies that make them retire old ones on a published schedule. When a model is retired, anything built on it has to move to a newer one. If nobody does that, the tool can stop working on a date nobody told you about.
This post explains what that means, lists the dated retirements as of November 2026, and gives you a simple routine to catch problems before your customers do. It's one reason AI is a job that needs keeping up, not software you install once.
What a model retirement is
Think of a model as the version of the AI brain your tool uses. Anthropic (Claude) and OpenAI (ChatGPT) both keep a list of versions, and they shut down older ones on set dates.
If you only use the chat app, they move you along for you. The risk is with anything built on top: a booking assistant, an automatic reply, a form that summarizes leads, a workflow that calls the AI behind the scenes. Those usually name a specific model. When that model is switched off, the call fails or gets redirected, and the result can be an error, a different style of answer, or a different answer altogether.
The vendors do give notice. Anthropic's docs say it gives at least 60 days' notice before retiring a publicly released model. But the notice goes to whoever holds the account, and that's often the person who built your automation, a vendor, or an employee who left.
The dates, as of November 2026
These come from each company's own deprecation pages. They change, so check the source before you rely on any date.
Anthropic, per its model deprecations page:
- Claude Opus 4 and Sonnet 4 were retired June 15, 2026.
- Claude Opus 4.1 was retired August 5, 2026.
- Claude Sonnet 4.5 was deprecated September 30, 2026 and retires November 30, 2026.
- The docs say deprecated models are likely less reliable and tell users to test new models well before the date.
OpenAI, per its deprecations page:
- The gpt-5 family and o3 shut down December 11, 2026.
- GPT-4o snapshots, GPT-4o mini, GPT-4.1 and GPT-4.1 mini shut down January 14, 2027.
- whisper-1 and the gpt-4o transcribe models shut down February 26, 2027.
- Some tooling is marked for shutdown November 30, 2026, including
/v1/prompts, Evals and Agent Builder.
We didn't read Google's Gemini deprecation page for this post, so we don't list Google's dates. If you use a Gemini-based tool, ask the person who set it up to check.
None of this means the sky is falling. Most owners are on the chat apps, and they'll be moved over without noticing. The point is the custom stuff. If someone built you an assistant in the last two years, it's worth one question: "Which model does it run on, and when is that retired?"
What breaks, and what the customer sees
Here are the realistic ways it goes wrong.
- It stops working. The tool throws an error. A website chat box goes quiet, or leads stop getting summarized. If the tool fails silently, you may not know for weeks.
- It works, but differently. A newer model may be more or less wordy, follow your rules differently, or handle your edge cases another way. The tone changes. It starts adding things you told it not to.
- It gets worse. This isn't only about retirements. A Stanford and Berkeley study found one model family scored 84% on one task in March 2023 and 51% on the same questions in June. That's old data on one task, and critics say part of the gap was how it was measured. But it supports the habit: hosted models change, so you recheck.
- The prompts and tools around it vanish. If your setup depends on a builder tool or saved prompt object that's retired, you lose the setup itself, not just the model.
Picture the customer's view. Someone texts your shop at 9 p.m. The assistant that used to say "We're open until 6, want me to book you in?" now returns nothing, or says something odd. They don't think "model retirement." They think you're closed or sloppy.
A retest routine you can run in an hour
You don't need to be technical. You need a short list and a calendar.
Once, when it's set up
- List every AI tool that talks to customers or touches customer data. Chat box, phone assistant, auto-replies, lead summaries, reminder texts.
- For each, write down who built it, who holds the account, and which model it uses. If nobody knows, that's your first finding.
- Write 10 to 20 test questions. Real things customers ask, including the awkward ones: a price question, an after-hours request, something it should refuse, a complaint. Add the correct answer for each.
- Save the results from today. That's your baseline.
Every month, and any time a model you use gets a retirement notice
- Look at the vendors' deprecation pages (links above) or ask whoever runs your setup to.
- Run your test questions. Read the answers. Compare to the baseline.
- Check the basics. Does it still hand off to a person? Does it still avoid quoting prices? Does it still log leads?
- Check the bill. A model swap can change usage costs. See how AI pricing works now.
- Write one line: what changed, what you tested, what you fixed. Keep it. That record is your proof if something goes wrong.
Before a retirement date
- Switch to the replacement model in a test copy first, if your setup allows it.
- Run the same 10 to 20 questions.
- Fix the instructions that behave differently.
- Switch over a few days before the date, not on it.
What to ask whoever built your AI
- Which model does each tool run on?
- Who gets the retirement notices, and what's your process when one arrives?
- How do you test after a switch?
- What does it cost when the model changes?
- Do I own the instructions, the test list and the accounts? (Gartner told buyers to ask who owns what the system learns, as reported by CIO.com.)
A good answer is specific. A vague one means nobody is watching. For a bigger checklist, see what a website and AI maintenance plan should include.
Model retirements are one reason upkeep is a step of its own in any setup. Our guide to AI implementation for small business shows where it fits.
Frequently asked questions
What is an AI model retirement?
It's when an AI company switches off an older version of its model on a published date. Tools built on that version have to be moved to a newer one. Anthropic's docs promise at least 60 days' notice. If nobody makes the move, the tool can fail or behave differently. What is managed AI covers who should be watching.
Will ChatGPT or Claude stop working for me?
If you use the chat apps, the company moves you to current models. The risk is custom tools built on a named model, like an assistant, an automation or a form that summarizes leads. Ask whoever built it which model it uses and when that model is retired.
How often should I retest my AI tools?
Monthly is a good habit, plus any time a model you use gets a retirement notice. Keep 10 to 20 real customer questions with the right answers, run them, and compare to your saved baseline. It takes about an hour once it's set up. A website maintenance plan can include it.
Does a retirement cost extra?
It can. Moving to a newer model takes testing time, and sometimes the new model's usage pricing differs. Ask your vendor in writing what's included in a model change before you sign. We cover bills in AI usage pricing explained.
How Di-Hy fits
Retesting when a model changes is part of what Keep It Running is for. We set up what you need, train your team, and keep an eye on the parts that change under you. If you already have a chatbot or automation built by someone else, a free audit will list what you have and which model each one uses. We're in Ybor City and work across Tampa Bay.
Next step: pick one AI tool you rely on and find out which model it runs on. If you can't find out, call us.
Dates come from Anthropic's and OpenAI's own pages as of November 15, 2026, and can change. Check the source before you act.


