Model aliases vs pinned snapshots

OpenAI

OpenAI model pages say snapshots let you lock a specific version so performance and behavior stay consistent. The same pages list the snapshots and aliases for that model.

The changelog records when a short slug is pointed at a new snapshot. On January 13, 2026 it moved the sora-2 slug to sora-2-2025-12-08, and it moved gpt-realtime-mini, gpt-audio-mini, gpt-4o-mini-tts, and gpt-4o-mini-transcribe onto 2025-12-15 snapshots. The previous snapshot keeps its dated id.

chat-latest points at the latest Instant model used in ChatGPT, and that page says the underlying snapshot will be regularly updated. The July 9, 2026 changelog entry says the gpt-5.6 alias routes requests to gpt-5.6-sol.

The API backwards-compatibility note says prompting behavior between snapshots can change. It says the way to keep prompting and output consistent is to use a pinned model version and to run evals.

Sources: GPT-5.6 Sol model page , Chat Latest , OpenAI changelog , OpenAI API backwards compatibility , checked October 20, 2026.

Anthropic

Anthropic says each Claude model ID is a pinned version for the life of that ID. That guarantee does not cover the convenience aliases on some earlier models.

Before the 4.6 generation, the Claude API also accepts a shorter alias such as claude-sonnet-4-5. That alias points at the most recent dated snapshot for that minor version. A 4.6-generation id such as claude-sonnet-4-6 is not an alias. It is the snapshot. Anthropic does not update the weights of an existing model ID. A newer model ships under a new ID.

Weights stay fixed for an ID. Serving pieces such as the request router, safety classifiers, and sampling logic can still change, and Anthropic says that can produce minor output differences on the same ID.

On Amazon Bedrock the older dated form looks like anthropic.claude-sonnet-4-5-20250929-v1:0. From Sonnet 4.6 on, Bedrock drops the -v1 suffix. Google Cloud uses an @ before the date on the older ids.

Sources: Claude model IDs and versioning , Claude models overview , checked October 20, 2026.

Gemini API

The Gemini models page, in the naming note dated September 2025, splits names into stable, preview, latest, and experimental. It says older models can use a different pattern, so read the exact string you call.

A stable name points at one stable model and usually does not change. The page's example is gemini-3.6-flash. A latest name, such as gemini-flash-latest, is an alias that is hot-swapped on every new release of that variation. Breaking changes get two weeks of email notice before the version behind latest changes.

A preview name can be used in production and is deprecated with at least two weeks of notice. Experimental names are not stable, and the endpoint can change.

Firebase's model list calls gemini-2.0-flash an auto-updated alias and records its shutdown on June 1, 2026. That page also says not to use -latest aliases, because they hot-swap and only breaking changes get the two-week email.

Sources: Gemini API models , Firebase AI Logic models , checked October 20, 2026.

Azure OpenAI

On Azure you call a deployment. The deployment has a model version and a versionUpgradeOption. A specific version stays selected until you change it or until that version's retirement date.

OnceNewDefaultVersionAvailable upgrades the deployment to the new default within two weeks of that designation. OnceCurrentVersionExpired upgrades when the current version reaches its retirement date. null means OnceCurrentVersionExpired. NoAutoUpgrade does not move. At retirement that deployment stops working.

Automatic updates are documented for Standard deployment types. Provisioned deployments are not auto-upgraded.

Sources: Working with Azure OpenAI models , Foundry Models lifecycle , checked October 20, 2026.

Amazon Bedrock and Vertex AI

Bedrock InvokeModel takes a modelId. That value is a foundation model ID, or an inference profile ID or ARN. ListFoundationModels returns the model ID. ListInferenceProfiles returns the profile ID. The lifecycle pages say Bedrock dates can differ from the model provider's dates, and that lifecycle state can differ by Region. They do not describe a short alias that is later pointed at a new snapshot.

The current Gemini Enterprise Agent Platform model versions page lists a model ID and a retirement date for that ID. It says a listed retirement date will not be moved earlier. It does not describe a latest alias. Use the Gemini API section above for gemini-flash-latest and the other latest names.

Sources: Bedrock inference API , Bedrock model lifecycle , Vertex model versions , checked October 20, 2026.

Questions