Your favorite AI model will be retired. Then what?

In April 2025, OpenAI removed GPT-4 from ChatGPT. The model a hundred million people had built habits around just stopped being an option, replaced by whatever came next. If you'd tuned prompts to it, built workflows on it, or simply liked how it wrote, none of that mattered. This is normal. Every cloud model you use today is on a deprecation schedule somewhere. The only models that never retire are the ones sitting on your own disk.

The retirement ledger so far

A short, public history: OpenAI retired GPT-4 from ChatGPT in April 2025 and has cycled GPT-3.5 and a string of dated API checkpoints out of service on published deprecation schedules. Anthropic has formally deprecated its early Claude families the same way. Both companies maintain deprecation pages because this is routine fleet management for them: old models cost serving capacity that new models use better.

Reasonable from their side. From yours, it means the tool you rely on has an expiry date you don't get to see in advance.

We have retired models too. Twice. Here is what it cost.

It would be easy to write this page as though retirement is something other companies do. It is not. This app has retired two tiers, and the interesting part is not that it happened — it is what the code has to do so that it does not strand the person who was using the retired thing.

A retirement is two changes, and only one of them is obvious

Dropping the catalog entry is the obvious half. The other half is a line in a table the user never sees, mapping the old id onto the new one. Skip it and a stored selection carrying the retired id reaches the model loader and gets 400 unknown model — a hard error, not a fallback.

Which means the people it breaks are precisely the people who were using the thing you retired. Everyone who had moved on already is fine. That is the shape of most migration bugs: the blast radius is exactly the users with the strongest claim on you.

The context number that looked like a loss and was not

Folding Code into Core took a tier with a 65,536-token default context and pointed it at one with 32,768. On its face that is a halving, and it is the kind of number a changelog gets shouted at over.

It was not a loss, and the reason is worth knowing because it is a real property of running models on your own machine rather than someone else's. The context default is raised from its base toward 262,144 depending on how much RAM is actually free, so above roughly 16 GB of headroom both tiers already resolved to the same window. Below it they diverged — and the retired tier took more, 4 GB of key-value cache against 2 GB, only because the gate never returns below a tier's declared base. Both tiers ask for 24 GB of RAM, and free memory after a 15.13 GB model sits well under that crossover.

So the retired tier's larger number was a bypass of the memory gate on the smallest machines that could run it at all — the hardware least able to spare 4 GB for a cache. Removing it removed a way to over-commit, not a capability.

Why 'the new model is better' doesn't fix it

Successors are usually stronger on benchmarks. They are also different. Different writing voice, different instruction-following quirks, different failure modes. The prompts you spent months refining behave differently overnight, and entire workflows built on one model's behavior quietly degrade. The recurring "did it get worse?" complaint threads after every silent model swap are this exact phenomenon: not weaker models, moved ones.

For API users it's sharper. A deprecated endpoint is a production incident with a countdown attached, and re-validating an app against a new model is real engineering work you didn't schedule.

The version of this that can't happen to you

An open-weight model is a stack of files. Outlier's tiers, Nano through Plus 397B, are exactly that: weights on your disk (published openly on HuggingFace), run by an engine on your Mac. Nobody can retire them, because there is no server to turn off. The model that writes a certain way today writes the same way in five years, unless you decide to swap it. Upgrades exist (new tiers ship and Pro seats include them), but they arrive as an option you choose, not a migration you're forced through.

What this means for how you choose tools

If your AI use is casual, retirements are a shrug: you adapt in an afternoon. If you've built anything durable on a model (a writing voice, a coding workflow, an app, a research process), the deprecation schedule is a risk you're carrying whether you priced it or not. The hedge is cheap: keep the workflows that matter on weights you control, and rent the cloud for the jobs where its ceiling genuinely matters. Stability where you need it, novelty where you want it.

Frequently asked questions