AI models often vanish from consumer apps like ChatGPT or Claude without being deleted. Companies like OpenAI and Anthropic move models through a lifecycle of deprecation and legacy status, often keeping them accessible through developer APIs or storing them in “cold storage” for potential future reuse.
How do AI companies retire models?
Major AI laboratories follow structured lifecycles to manage model transitions. According to Anthropic’s documented process, a model moves from Active (fully supported) to Legacy (no updates), then to Deprecated (not recommended with a set retirement date), and finally to Retired, where all requests fail.
OpenAI utilizes a similar distinction. The company separates “legacy” models, which simply stop receiving updates, from “deprecated” models that have an official shutdown date scheduled. Google follows a different nomenclature, treating “deprecation” as the formal announcement and “shutdown” as the specific moment the endpoint is deactivated.
Why do models disappear from apps but stay in APIs?
A model’s disappearance from a user interface does not always indicate its death. When OpenAI removed several models, including GPT-4o and GPT-5, from the ChatGPT interface in February 2026, the company noted that no changes were made to the API at that time.
This distinction allows developers to continue building tools on stable, older architectures while the company pushes newer, more efficient models to general consumers. This creates a split experience: the consumer app receives the latest “flashy” updates, while the API serves as a stable environment for enterprise software.
In some cases, consumer backlash can reverse these decisions. OpenAI previously re-introduced a subset of GPT-4o to Plus and Pro users after reports from Reddit and other forums indicated users preferred its specific conversational style for creative tasks.
What happens to “retired” AI weights?
Some companies treat retired models as digital assets rather than obsolete software. Anthropic has publicly committed to preserving the weights of its released models. This policy acts as a form of “cold storage,” where discontinued models are kept in a digital warehouse for potential future use.
This approach is currently being tested as Anthropic prepares to retire its original Claude Opus 4 and Sonnet 4 models in mid-June 2026. The company has also indicated a philosophy of exploring “retirement interviews” to honor the specific preferences or behaviors expressed by a model before it is taken offline.
Beyond storage, models are often “recycled” through a process called distillation. Engineers take the intelligence of a large, older model and use it to train a smaller, faster, and cheaper successor. This effectively turns the “brain” of an old model into the foundation for the next generation.
Which AI models can never be shut down?
Open-weight models represent a category of AI that is immune to corporate retirement. Once companies like Meta, Mistral, DeepSeek, or Alibaba release the weights for models like Llama or Qwen, no single entity can switch them off.
These files live indefinitely on community hubs like Hugging Face. Because they can be “quantized”—a process that shrinks the model to run on laptops or phones—they continue to evolve through community fine-tuning long after the original creator has moved on to new projects.
Comparison of Model Retirement Policies
| Company | Primary Retirement Method | Key Characteristic |
|---|---|---|
| Anthropic | Cold Storage | Preserves weights for potential future revival. |
| OpenAI | App/API Split | Removes from consumer apps while maintaining API access. |
| Meta/Mistral | Open-Weight Release | Models exist indefinitely via community hosting. |
Frequently Asked Questions
How much notice do I get before an AI model is retired?
Notice periods vary by provider. Anthropic states they provide at least 60 days’ warning before retiring a publicly released model. Google publishes shutdown dates in advance, though consumer app changes may occur more rapidly.
Can I still use a model if it is “deprecated”?
Yes. A deprecated model still functions, but it is no longer the recommended option for new projects. It is essentially in a “waiting room” before its official shutdown date.
What is the biggest risk of model retirement?
The primary casualty is customization. When a base model is retired, any custom fine-tuning built on top of that specific model often becomes unusable, forcing developers to retrain their systems from scratch.
What do you think about the “retirement” of AI models? Do you prefer the latest updates, or do you miss the specific “personality” of older versions? Let us know in the comments below or subscribe to our newsletter for more AI industry insights.
