AI Model Shutdowns: What to Do If Your App Runs on GPT-4o, GPT-5 or Claude Sonnet 4.5

If your product has an AI feature, there's a fair chance it's built on a model with an end date. Between 30 November 2026 and 26 February 2027, OpenAI and Anthropic will shut down several of the models that small businesses and startups built on in 2024 and 2025. That includes GPT-4o, GPT-4.1, GPT-5, o3, Whisper and Claude Sonnet 4.5.

When a model is retired, requests to it simply fail. Anthropic's documentation says this plainly: "Requests to models past the retirement date will fail." Your chatbot, summariser or voice agent doesn't get worse. It stops.

The short answer

Find out which model IDs your app calls, check them against the dates below, and switch to the recommended replacement before the date. Test the new model on real examples first: a replacement model is a different model and can behave differently. Allow a few days for a simple feature and a couple of weeks for anything customers depend on.

Which models are shutting down, and when?

These are the dates from the official deprecation pages as of 31 October 2026 (OpenAI deprecations, Anthropic model deprecations).

Date	What shuts down	Recommended replacement
30 Nov 2026	Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)	claude-sonnet-5-5
30 Nov 2026	OpenAI Agent Builder, the reusable prompts API (v1/prompts) and the Evals platform	Agents SDK; prompts kept in your own code; Promptfoo for evals
11 Dec 2026	gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-pro	gpt-5.6-sol, gpt-5.6-terra or gpt-5.6-luna, depending on the model
6 Jan 2027	tts-1, tts-1-hd, gpt-4o-mini-tts (text-to-speech)	gpt-realtime-2.1-mini
14 Jan 2027	gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini	gpt-6-luna
20 Jan 2027	gpt-realtime, gpt-4o-realtime, gpt-audio and their mini versions (voice)	gpt-realtime-2.1, gpt-realtime-2.1-mini, gpt-audio-1.5
26 Feb 2027	whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe	gpt-live-transcribe or gpt-transcribe

Two things have already happened. OpenAI's Assistants API shut down on 26 August 2026, replaced by the Responses and Conversations APIs. If your chatbot was built on Assistants and hasn't been touched since, it's already broken. Several older OpenAI models, including o1, o4-mini and gpt-4-turbo, also shut down on 23 October 2026.

One caveat: these dates apply to the providers' own APIs. If you use Claude through Amazon Bedrock or Google Cloud, Anthropic notes that those platforms "set their own retirement schedules", so check your cloud provider's model table as well.

How do I find out which model my app uses?

You don't need to read the code to get a first answer.

1. Check the provider's usage page. Anthropic's Console has a usage export that breaks usage down by API key and model. Look at your OpenAI usage dashboard as well. If a model on the shutdown list appears there, you have work to do.
2. Ask whoever built it for every place a model name appears: environment variables, config files, database settings. Model names are often hard-coded in more than one place.
3. Don't forget the side features. The main chatbot might be on a current model while transcription runs on whisper-1 and voice replies use tts-1. Voice apps usually touch three or four models.
4. Check for aliases. A setting like gpt-5 points at a specific model, and that model is on the list. Pinning to a name doesn't protect you.

Is it just a one-line change?

Sometimes. Changing the model name takes a minute. The work is in making sure nothing else changes with it:

• Different answers. A newer model may be more verbose, more cautious or better at following some instructions and worse at others. Prompts that relied on one model's habits often need adjusting.
• Broken formats. If your app expects JSON back, or a reply in a fixed shape, test that first. That's where migrations most often fail quietly.
• Removed settings. Anthropic's API version of 21 October 2026 rejects temperature, top_p and top_k with an error (migration notes). Code that sets them will fail on newer versions even if the model name is right.
• Different costs and speed. Price per token and response time change between models. Check both before you switch, not after the first invoice.
• Voice is the most sensitive. A new realtime model can change voice, pacing and how it handles interruptions. For a voice agent that talks to customers, that's a product change, not a maintenance task.

A safe migration checklist

This is the order I'd work in for a client's app:

1. List every model and API the app calls, with the shutdown date for each.
2. Collect 30–50 real examples of what the feature handles today: real questions, real documents, real calls (anonymised).
3. Run those examples through the old and new model side by side and compare the outputs. This is the step people skip, and the one that catches problems.
4. Fix prompts and parsing where the new model behaves differently.
5. Switch behind a setting, so you can roll back in minutes if something goes wrong.
6. Watch errors and costs for the first week.
7. Put the next date in the calendar. The replacement models have end dates too. Anthropic lists a "not sooner than" retirement date for every active model.

Step 3 is the reason to start now rather than in the last week. If the comparison shows the new model handles your use case worse, you want time to adjust prompts or try a different model, not a deadline.

What does it cost to migrate?

It depends on how many models the app uses and how much the output matters. These are the planning ranges I use before looking at the code. They're not quotes:

Situation	Typical effort
One chatbot or text feature, one model, simple output	A day or two, including testing
Several features, structured JSON output, or the Assistants API	One to two weeks
Voice agent (speech-to-text, model, text-to-speech, realtime)	Two weeks or more, mostly testing call quality

If you want a number for your own app, my cost calculator gives a starting estimate, and I'm happy to look at the specifics.

Should I switch providers while I'm at it?

Only if you have a reason. A forced migration is a good moment to compare, because you're testing a new model anyway, and running the same examples through a second provider adds little work. But switching providers also means new SDKs, new rate limits and new billing, so do it when it improves quality, cost or reliability, not by default.

What I do recommend is keeping the model name in configuration rather than code, and keeping your test examples. Then the next shutdown notice is a small task, not a scramble.

I've built AI features on these APIs for clients, including voice agents for loan and mortgage enquiries and the GPT-based job-application assistant Jackfruit.ai. If you're weighing up a voice agent rather than maintaining one, my notes on where AI voice agents work and where they fail are a good place to start.

FAQ

What happens if I miss the shutdown date?

Requests to a retired model fail with an error. Whatever feature depends on it stops working until the code points to a model that's still available. There's no slow decline first.

Will the provider warn me?

Anthropic says it notifies customers with active deployments at least 60 days before a public model retires. But notices go to whoever owns the API account, which isn't always the business owner. Check the deprecation pages yourself.

Can I just switch to the newest model and skip testing?

You can, but it's a gamble. Newer models are generally better, but "better" on average can still mean different output for your prompts. Testing a few dozen real examples is cheap insurance.

I'm not technical. Can someone just handle this?

Yes. A model migration is a well-defined job: audit, test, switch, monitor. If you'd like me to check your app, send me the details through my AI development page or contact me directly.