These pages define the words providers use when a model is on the way out. The first paragraph of each page is the definition. The sections under it say where OpenAI, Anthropic, Gemini, Vertex, Bedrock, and Azure use a different word. Example rows come from this site's model records. Last reviewed . Publisher: SunsetDiff.
Dates for a specific id stay on the model pages, the deprecated and retired lists, and upcoming shutdowns. Practical steps stay on the guides.
| Term | Definition |
|---|---|
| Deprecated model | A deprecated model is one the provider still serves after announcing that support will end. It is not retired yet: requests keep working until the shutdown date, unless that provider's page says a narrower rule. |
| Retired model | A retired model is no longer available, and requests to that id fail. Providers also call this moment shut down, sunset, or end of life, and they do not all name the same HTTP status. |
| Shutdown date | A shutdown date is the day the provider says the model stops being served. A date that only says not sooner than is a floor, not that day, and a first-party date does not move the date on Azure, Bedrock, or Vertex. |
| Sunset | On OpenAI's deprecations page, sunset and shut down mean the same thing: the model or endpoint is no longer accessible. Anthropic, Gemini, Vertex, and Azure say retirement or shutdown, and Amazon Bedrock says end of life. |
| Legacy model | A legacy model still runs, and the provider has told you to plan a move. OpenAI and Anthropic mean no more updates, Bedrock means the scheduled-retirement window, and Azure means an optional stage before Deprecated. |
| End of life | End of life (EOL) is Amazon Bedrock's name for the date a model is removed and requests to it fail. OpenAI calls that moment shut down or sunset, and Anthropic, Gemini, Vertex, and Azure call it retirement or shutdown. |
| Model snapshot | A model snapshot is a dated model id that stays on one version until that id is retired. A short alias can be pointed at a different snapshot later, and the dated id is the one that does not move. |
| Model alias | A model alias is a short name that can be pointed at a newer snapshot later, while a dated or pinned id stays put. OpenAI, Anthropic, and Gemini each publish an alias rule, and Amazon Bedrock's lifecycle pages do not. |
| Pinned version | A pinned version is a model id that stays on one snapshot until you change it or that id is retired. Put the dated id, or the Azure deployment version you selected, in production after you test it. |
| Preview model | A preview model is a pre-release id with a shorter life than a generally available model. OpenAI identifies them by the word preview in the name, and Gemini experimental names are a separate label for models that are not stable. |
| Auto-upgrade | An auto-upgrade is the host moving your deployment onto a newer model version without you changing the model string. Azure documents this for Standard deployment types, provisioned deployments are excluded, and a Gemini latest alias is a different swap. |
| Migration | A migration is the change from the model id you call today to the replacement id, including request fields the new id rejects. It is a change you make, unless the host documents an automatic upgrade. |
| Replacement model | A replacement model is the id the provider names as the substitute for the one that is going away. If the official page does not name one, there is no replacement to invent. |
| Notice period | A notice period is the minimum time a provider says it will give before retirement, and it is not the shutdown date. The date on the deprecation table is the day to migrate by. |
| Provisioned throughput | Provisioned throughput is reserved capacity for a deployment, as opposed to pay-as-you-go Standard. When a model retires, Azure does not auto-upgrade provisioned deployments, and Bedrock blocks new Provisioned Throughput for a model in Legacy. |
| Regional availability | Regional availability means the same model id can be active in one region and legacy or end of life in another. The date on the first-party API page does not move the date on Azure, Bedrock, or Vertex. |
| API version header | An API version header tells the provider which request contract to apply. On the Claude API the header is anthropic-version, and changing it can reject fields on a model that is still available. |
| Breaking change | A breaking change is a request setting or endpoint that fails after you switch model ids or move to a new API version. The new model can still be up, and the old field is what fails. |
| Fallback model | A fallback model is a second id a client tries when the first call returns an error. LiteLLM and OpenRouter document that list, and provider deprecation pages do not switch your request to the replacement. |
Deprecated means the provider still serves the model and has announced that it will stop. Retired means the model is no longer available and requests fail. The words are not identical on every provider. See Deprecated and Retired.
A snapshot is a dated id that stays on one version until that id is retired. An alias is a short name that can be pointed at a newer snapshot later. See Model snapshot and Model alias.
On OpenAI's deprecations page, legacy means the model or endpoint no longer receives updates and is expected to be deprecated later. Deprecated means retirement has been announced and a shut down date exists. See Legacy model.
It is the day the provider says the model stops being served. A date that only says not sooner than is a floor, not that day. See Shutdown date.