AI Model Deprecation: Retirement With No Handover

AI model deprecation is when a provider marks a model as no longer recommended, sets a retirement date, and then switches it off. Past that date, requests to it fail.

OpenAI shuts down the Assistants API on August 26, 2026. Anthropic retired Claude Opus 4.1 on August 5. Google switched off 3 Veo video models on June 30 after announcing it on June 15, so 15 days of warning.

That is one season. None of it counts as unusual now.

We watch this more closely than most at ChatFuse, because routing across more than 100 models means every one of these notices reaches us first. Keeping the list current is ordinary maintenance here.

Spend months teaching an AI how your business works and you pick up an assumption without noticing it, that the effort is accumulating somewhere permanent. It isn't. The model gets switched off on a date somebody else chose, and whatever replaces it knows none of what you taught the last one.

What is AI model deprecation?

AI model deprecation is a lifecycle stage, and every major provider runs one now. Anthropic publishes 4 states: active, legacy, deprecated, retired. A deprecated model still answers, but it is no longer recommended and it carries a retirement date. A retired model returns an error.

The vocabulary is calm on purpose. It reads like a version bump. What it describes is a capability you built on being withdrawn, on a schedule you had no say in.

Which AI models are retiring in 2026?

Quite a few, and the pace went up this year. The dates below come from the providers' own documentation rather than the trackers, because a lot of the trackers have them wrong. ChatFuse checked every one of them against the source page on August 21, 2026.

Worth saying what is not on that list. Plenty of articles have GPT-4o retired back in April. OpenAI's deprecation page gives GPT-4o no retirement date at all. Others print fixed October dates for the Gemini 2.5 cluster, which Google has not announced. If you are planning against dates you read in a roundup, go and check them.

How much notice do AI providers actually give?

Anywhere from about a year down to about 2 weeks. Anthropic commits publicly to at least 60 days for publicly released models and has stayed close to it. OpenAI gave the Assistants API a full year. Google gave 3 Veo models 15 days.

That footnote is doing a lot of work. Google's page says the dates it publishes are the earliest a model might be retired, and that the real date gets communicated later. So with one of the 3 biggest providers, the schedule you can see is not a thing you can firmly plan against.

Why is it worse when the tooling retires?

Because swapping a model is a configuration change and losing a tool is a rebuild. Anyone who built on the OpenAI Assistants API is not editing a model name this month. They are moving onto a different API with different concepts.

This is the part ChatFuse absorbs on your behalf, and it is the part that costs real engineering time. Look at what OpenAI announced on June 3 alone. Agent Builder, the Evals platform, and reusable prompts, all closing November 30. Those are the tools teams used to build agents, test them, and keep their prompts in one place. Adopt all 3 and you rebuild all 3, on a date somebody set for you.

What actually breaks when a model retires?

The context breaks. The model itself is replaceable in an afternoon. What is not replaceable is everything the old one had absorbed about how your business works, because none of that transfers to whatever comes next.

Which is where the word retirement turns out to be exact. A person retiring spends their last weeks handing over. They write things down. They train somebody. Companies treat that as a known risk, give it a name, and put a budget against it.

Anthropic has thought about this in terms that are hard to read as anything other than the human version. Its deprecation page lists model welfare among the downsides of retiring models, commits to preserving model weights long term, and says the company hopes to make past models publicly available again one day. A pension and an alumni network, written into API documentation.

Do companies plan for model retirement?

Mostly not, although the market has hedged without calling it that. LangChain's State of Agent Engineering surveyed 1,340 practitioners between November 18 and December 2, 2025, and found more than 75% already running several models in production or development. Very few would call that a deprecation strategy. It is one anyway.

Businesses run succession planning for people and nothing equivalent for models, even where a model is doing work a person used to do. Nobody owns the question. So it turns up as an outage.

What survives a model retirement?

Only what you hold outside the provider. If your context, preferences and working history sit inside one vendor's product, they retire on that vendor's schedule. Keep them somewhere you control and the model underneath can change while the work carries on.

That is the argument for spreading across models instead of committing to one, and it is what we built ChatFuse around. The ChatFuse Orchestrator routes every prompt across more than 100 models from OpenAI, Anthropic, Google, Meta and others, so a retirement changes which model answers instead of starting a migration. Persistent memory lives in your ChatFuse account, which means what the system knows about your work is not sitting inside the model that gets switched off. If you are thinking about AI as a set of roles rather than one tool, our AI org chart post covers how we structure that.

I should be straight about the limit. We do not escape any of this. ChatFuse runs on the same providers, so when Anthropic retires a model we do that migration too. The difference is who absorbs it. We take on keeping more than 100 models current, and your memory sits somewhere a retirement cannot reach.

Frequently asked questions

What does deprecated mean for an AI model?

Deprecated means the model still works but is no longer recommended, and the provider has given it a retirement date. Once that date passes the model moves to retired and requests to it fail.

How much notice does Anthropic give before retiring a model?

Anthropic commits to at least 60 days of notice before retiring a publicly released model, and emails customers with active deployments. Claude Opus 4 and Claude Sonnet 4 were deprecated on April 14, 2026 and retired on June 15, which is 62 days.

Is the OpenAI Assistants API being shut down?

Yes. The OpenAI Assistants API shuts down on August 26, 2026, about a year after it was deprecated in August 2025. Teams still on it have to move to the Responses API or rebuild.

Was GPT-4o retired in 2026?

No. Several articles claim GPT-4o was retired in April 2026, but OpenAI's own deprecation page lists no retirement date for it. Check any deprecation claim against the provider's documentation before you plan around it.

How do I protect my business from AI model deprecation?

Keep your context somewhere the provider does not control, and avoid building a critical process on a single model. Running across several models means a retirement changes which model answers rather than stopping the work.

Does using multiple AI models cost more?

Not necessarily. ChatFuse includes more than 100 models in one subscription and routes each prompt to a model that suits the task, so a smaller and cheaper model often handles work that would otherwise go to the most expensive one. See pricing for what that covers.

Model retirement is a scheduled event now, and the providers have published the dates. The question worth asking is whether the thing your business depends on lives somewhere those dates can reach.

Start free with ChatFuse and keep your memory somewhere the retirements cannot.