AI Model Deprecation: Retirement With No Handover

AI model deprecation is when a company says a model isn't recommended anymore, picks a date to turn it off, and then pulls the plug. After that, any call to it just fails.

OpenAI is shutting down its Assistants API on August 26, 2026. Claude Opus 4.1 from Anthropic got retired on August 5. Google turned off 3 of its Veo video models on June 30. They'd only given 15 days' notice, announcing it back on June 15.

That's just one stretch of time. Nobody even blinks at this anymore.

We track these changes more than other companies, because with more than 100 models in ChatFuse, every one of these announcements hits us before anyone else. Just keeping the roster up to date is a normal part of our day.

It's easy to think the work you put into teaching an AI how your company operates is building up into something lasting. But it doesn't. They shut the model down whenever they decide to, and whatever new one comes next won't know a thing you showed the old one.

What are the stages of a model's retirement?

Every AI company keeps a schedule for its models now. At Anthropic, there's 4 official labels: active, legacy, deprecated, and retired. A deprecated one will still take your requests, but they don't suggest using it anymore and they've already set a date when it'll stop completely. Once it's retired, all it does is give you an error.

The words sound neutral, almost like a routine update. But what's really happening is a feature you built something on is getting taken away, and you didn't get to pick the timeline.

Which AI models are retiring in 2026?

Quite a few models got pulled this year, actually, and the pace really picked up. We took the dates straight from the providers' own pages instead of those tracker sites, because many of the trackers have bad info. ChatFuse verified each one against the original source on August 21, 2026.

Plenty of articles claim GPT-4o got retired in April. That's not on our list because OpenAI's own deprecation page doesn't show an end date for it. You'll also see some reports putting Gemini 2.5 on a fixed October schedule, but Google hasn't announced that. If you're making plans based on dates you read somewhere, go check the source yourself.

How much notice do AI providers actually give?

It can be anywhere from a year down to just 2 weeks. Anthropic says they'll give you at least 60 days for any public model, and so far they've stuck to that. The Assistants API from OpenAI had a full year. Google told people they had 15 days for 3 different Veo models.

That footnote carries some weight. Google's own page says the dates they list are just the soonest a model might get retired, the actual shutdown date comes later. So for one of the 3 biggest players, what you see isn't something you can really schedule around. Here are the cases where both announcement and shutdown dates were public.

Why is it worse when the tooling retires?

Because changing a model is just a settings edit, but losing a feature means rebuilding everything from scratch. Anyone who built on the Assistants API isn't just picking a different model name now. They have to move their entire project to a whole different system that doesn't even work the same way.

ChatFuse handles that part for you, and that's the part that actually burns through engineering hours. Just look at what got turned off from the June 3rd announcement. Agent Builder, the Evals platform, and the place for reusable prompts are all gone after November 30. Those were the tools teams relied on to create agents, to test them, and to manage their prompts. If you used all 3, you have to rebuild all 3, and you have to do it by a deadline you didn't choose.

What actually breaks when a model retires?

When a model goes away, you can swap it out in a day. But you can't replace everything it learned about how you run your business. That knowledge doesn't move over to whatever you use next.

That's why we call it a retirement. A person who retires spends their final weeks on handover. They document what they know. They train a replacement. Companies see that as a standard risk, so they name it and set aside money for it.

Anthropic's approach looks downright human. The deprecation page they keep actually lists model welfare as a downside when they retire old models. They commit to preserving the model weights for the long haul and say they want to make old models public again someday. It's like a pension plan and an alumni network, written right into the API docs.

Do companies plan for model retirement?

Mostly not, but the market's already doing it without the name. LangChain asked 1,340 people from November 18 to December 2, 2025, and more than 75% said they use several models right now. Hardly any of them would say that's a plan for deprecation. But it works that way.

Companies plan for when people leave, but they don't do that for models, even if a model does a job a person once did. Nobody's in charge of that question, which is the same problem we talk about in who owns AI in your company. What you get instead is an outage.

What survives a model retirement?

You only own what you keep outside. If your history, context, and preferences live inside one company's product, they'll disappear when that company decides it's time. Keep them in a place you control, and you can swap the model underneath without interrupting your work. That's the core idea behind own your AI context, rent the models.

This is why using multiple models beats sticking with just one, and it's the foundation of ChatFuse. Our Orchestrator directs every prompt across more than 100 models, sourced from Google, Meta, OpenAI, Anthropic, and others, so a model retirement just means a different model picks up the task, it doesn't mean starting over. Your memory stays persistent in your ChatFuse account, so everything the system learns about your work isn't trapped inside a model that might get shut down. If you see AI as a team of roles instead of a single tool, our post on the AI org chart explains how we set that up.

I need to be clear about one limit. We aren't immune to this. ChatFuse uses those same providers, so if Anthropic retires a model, we have to handle that migration too. The real difference is who deals with it. We handle the job of keeping all those models up to date, and your memory stays safe in a place those retirements can't touch.

What does deprecated mean for an AI model?

The model still works for now, but the company that made it says not to start new projects with it. They set a date when it'll stop completely. After that, calls to it just fail. Take Anthropic, they list 4 stages for a model: active, legacy, deprecated, and retired.

How much notice does Anthropic give before retiring a model?

Anthropic gives people at least 60 days' warning before they take a public model offline. They also email anyone who's actively using that model. Claude Sonnet 4 and Claude Opus 4 were deprecated on April 14, 2026. Their retirement happened on June 15, which works out to 62 days of notice.

Is the OpenAI Assistants API being shut down?

Yes. It gets shut down on August 26, 2026. That's about a year after it went into deprecation back in August 2025. If you're still using it, you need to move over to the Responses API or build something new.

Was GPT-4o retired in 2026?

No. A few posts said GPT-4o got retired in April 2026, but if you check OpenAI's own deprecation page, there's no end date listed for it. Always look at the source docs before you make plans based on a claim like that.

How do I protect my business from AI model deprecation?

Don't keep your context where the vendor controls it. Don't build a key process around just one model. If you run across more than one, a model retirement just changes which one answers. It doesn't stop the work. You can also change the AI model during a conversation if one gets pulled.

Does using multiple AI models cost more?

Not necessarily. With ChatFuse, you get a subscription that covers more than 100 models. We route your prompt to the one that's best for the job. A lot of the time, a smaller, cheaper model can handle the work just fine instead of the most expensive one. Check pricing to see what's included.

Where to start

You can't treat AI model retirements like a surprise anymore. Providers set the schedule and publish the dates. So the real issue is whether the tools you rely on are sitting where those dates can find them.

Start with ChatFuse for free and put your memory somewhere those retirements won't touch.