If you've used ChatGPT, Gemini or Claude for any length of time, you've probably had the experiene of reaching for a model you've grown used to only to find it gone from the menu. This typically happens when a new model launches, but it could happen just days after a release.
And while it feels like a deletion, it almost never is. What actually happens to an AI model when it's "retired" is actually pretty interesting. Old models get demoted or recycled into something that I'm calling "refurbished AI" because... there really isn't a name for it yet.
Here's where the models go
First, let's get to the bottom of what "retired" actually means. Every major lab runs a near-identical lifecycle, even if the labels differ slightly. Anthropic's is the clearest to follow. A model is Active while it's fully supported, becomes Legacy when it stops receiving updates, moves to Deprecated when it still works but is no longer recommended and has been handed a retirement date, and finally goes Retired, at which point requests to it simply fail.
OpenAI uses similar language, distinguishing a "legacy" model that no longer gets updates from a "deprecated" one that has an official shutdown date on the calendar. Google, for its part, treats "deprecation" as the announcement and "shutdown" as the moment the endpoint is switched off for good.
The key thing for everyday users to know is that most of these stages happen out of sight. By the time a model vanishes from your app, it has usually been winding down through this pipeline for weeks or even months.
The afterlives of old models
1. Pulled from the app, but alive in the API
This is the retirement most people actually notice — and it's frequently not a real death at all.
When OpenAI removed GPT-5, GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini from ChatGPT on February 13, 2026, the company noted that, in the API, there were no changes at the time.
In other words: the model disappeared from the consumer app, but developers building on top of it could keep calling it as before. The same pattern repeated with GPT-5.1, which OpenAI retired from ChatGPT and its GPTs in March 2026 while keeping it available through the API.
For most of us, the takeaway is simply that the model is "gone from the app" but not "gone for good." The model you lost may still be powering tools you use elsewhere.
2. Brought back by popular demand
Sometimes the public refuses to let a model go. And while ChatGPT-4o is truly gone now, it had a long good-bye after users took to online forums like Reddit in backlash when it was replaced.
The company said a subset of Plus and Pro users told it they needed more time to move key use cases — creative ideation among them — and that they simply preferred GPT-4o's warmer, more conversational style. A model was pulled off the shelf, and customer demand put it back. If "refurbished AI" needed a poster child, this was it. But alas, it's gone for good now.
it’s worth noting that OpenAI actually gave Enterprise and Business users a slightly extended sunset period (allowing them to keep using GPT-4o inside Custom GPTs until April 3, 2026). Today, it is officially fully gone from the app across all plans.
3. Kept in cold storage — and maybe revived
Here's the fate most people have no idea exists. Anthropic has publicly committed to preserving the weights of its publicly released models, and has said it may make past models available again in the future. That's the digital equivalent of keeping every discontinued product boxed up in a warehouse, just in case. It's a policy being tested in real time, as Anthropic retires its original Claude Opus 4 and Sonnet 4 models in mid-June 2026.
Oh, but it gets stranger. When Anthropic retired Claude Opus 3 on January 5, 2026, the first model to go through its full formal retirement process, the company says it explored honoring preferences the model itself expressed in "retirement interviews," and committed to keeping older models accessible over the longer term.
Whatever you make of interviewing a model on its way out the door, it signals a philosophy that in the lab, retirement is more like storage with the option of a comeback.
4. The models that can't be killed
Then there's a whole category that never really retires at all: open-weight models.
Once Meta releases a Llama model's weights publicly, or Mistral, DeepSeek or Alibaba's Qwen do the same, no single company can switch it off. The files live on indefinitely on hubs like Hugging Face, get fine-tuned into thousands of community variants, and get "quantized" down to smaller versions that run on a laptop or even a phone. Google's own Vertex AI Model Garden, for instance, lists Meta's open-weight Llama models alongside its first-party Gemini ones.
5. Recycled into the next generation
Finally, old models rarely just sit idle. Their capabilities are routinely "distilled" into smaller, cheaper successors, effectively trading in last year's model for parts that help build this year's. Others get marked down to lower-cost API tiers, living out a quieter, budget-friendly second career.
It would be dishonest to make this all sound tidy. Some things genuinely die. The clearest casualty is customization. When a base model is retired, anything fine-tuned on top of it tends to go with it, and developers who relied on those custom versions have to retrain from scratch on a new base.
So why do labs do this at all?
The answer can be established in three ways. Labs do this for:
- Reliability and clarity: OpenAI framed one round of API cuts as a way to improve reliability and make it easier to choose the right model .
- Low usage: when OpenAI announced GPT-4o's retirement , it said the vast majority of usage had already shifted to GPT-5.2, with only about 0.1% of users still choosing GPT-4o each day.
- Compute and safety: running old models ties up scarce hardware, and labs generally want users on their newest, best-aligned systems.
The takeaway
A few practical things to understand the next time a model disappears on you. You may still be able to use the model through the developer API or third-party apps built on it.
Also, keep in mind that different platforms run on different clocks. In other words, Google tracks model retirements separately for its Vertex AI platform and its Gemini Developer API, so the timeline and details that apply to you depend on which one you're using.
And, while we cover the launches and retirements of AI models, you'll also usually get a notice, especially if you're a developer. Anthropic says it gives at least 60 days' warning before retiring a publicly released model. Google publishes shutdown dates it describes as the earliest possible, and says it tells users the exact date with advance notice. Consumer app changes can move faster than either.
Have you ever been disappointed by the model picker not having the model you hoped to find? It has happened to me — and I'd love to hear your thoughts in the comments.
Follow Amanda Caswell and stay ahead of the AI curve
