A floating model alias is convenient for exploration because it lets the provider change what sits behind a familiar name. That is exactly why it is a weak production release boundary. If behavior changes while application code and configuration remain untouched, rollback and incident analysis become guesswork.
CometAPI’s GPT-5 access guide illustrates the distinction. The dated article lists floating names such as gpt-5 alongside a dated identifier such as gpt-5-2025-08-07, then explicitly says its latest listed versions are only current as of publication. It also recommends tracking alias usage and checking the exact supported model name when access fails. The article does not establish which identifiers are available today.
Separate discovery from release configuration
Use floating aliases in a discovery environment to find candidate updates. Use an immutable versioned identifier in production when the provider exposes one. Keep both values in the release record:
task: support-summary
candidate_alias: <floating alias>
deployed_model_id: <verified versioned id>
contract_revision: 7
The alias explains the intended family. The deployed ID explains the actual release. Business code should reference the task name rather than either model string directly.
If a provider exposes only a floating alias, record the catalog snapshot time and the model identity returned by each request. That improves observability but does not create immutability. Treat a later behavioral difference as possible upstream drift, not automatically as an application regression.
Promote a model like a dependency
A release process can stay small:
- Resolve the alias against the current catalog for the intended account and endpoint.
- Run contract fixtures for output schema, tool calls, safety behavior, token limits, latency, and cost boundaries that matter to the task.
- Compare the candidate with the deployed version on the same fixed inputs.
- Canary a bounded portion of eligible traffic.
- Promote the exact versioned ID and retain the previous one as the rollback target.
Do not declare success because a request returned HTTP 200. Record semantic acceptance and the returned provider model identity. A gateway can accept the outer request while the chosen route rejects a field, returns an unexpected shape, or behaves differently on a critical prompt.
Fail closed on missing versions
At startup, validate every configured production ID against the current route and account. A missing version should stop that task or move it to a separately tested fallback. Silently substituting the floating alias defeats the release boundary.
Fallbacks need versioned identities too. Each one must satisfy the same task contract, and the system should log whether the primary or fallback served the result. A version retirement then becomes a planned migration: select a candidate, test it, authorize the configuration change, and preserve a rollback path.
Cache discovery results only as candidate metadata. Deployment configuration should not change merely because a background catalog refresh sees a newer alias target. That separation allows teams to observe upstream movement immediately while choosing a controlled maintenance window for contract tests and promotion.
Versioned IDs do not prevent model drift inside a provider’s implementation, but they remove one large ambiguity. They make “what changed?” answerable enough to operate the application deliberately.