Every hosted model has an end date, even if nobody has announced it yet. A deprecation notice is the start of the clock.
How it works
Providers publish a lifecycle with named stages. Anthropic’s is the most detailed:
- Active: fully supported and recommended.
- Legacy: “will no longer receive updates and might be deprecated in the future.”
- Deprecated: “still functional but no longer recommended”, with a replacement and a retirement date.
- Retired: “no longer available for use. Requests to retired models fail.”
OpenAI uses two terms: deprecated, meaning “OpenAI has announced the model’s retirement”, and shutdown, when “The model no longer accepts requests through the standard API.”
Notice periods, as published on 15 November 2026:
| Provider | Stated notice | How customers hear |
|---|---|---|
| OpenAI | At least 6 months for generally available models; at least 3 months for specialised and image generation models; preview models may go sooner | Email and the deprecations page |
| Anthropic | At least 60 days for publicly released models | Email and the documentation |
| Google (Vertex AI) | Standard models “available for at least 12 months after release”; short-term models get a retirement date that gives “at least 45 days to migrate”; dates “won’t be moved to an earlier date” | Fixed dates in a published table |
A real example: Anthropic notified Claude Sonnet 4.5 users on 30 September 2026 of retirement on 30 November 2026, with claude-sonnet-5-5 as the replacement.
Why it matters
- Silent breakage. After the date, a hard-coded model ID returns errors. Keep model IDs in config, not code.
- Behaviour changes. A replacement model answers differently. Anthropic advises testing “well before the retirement date”. Re-run your eval on the replacement, and see evaluate an LLM before switching.
- Cost and limits change too. The replacement can have different prices, context length and rate limits.
- Audit usage. Find every service and key that still calls the old model before the deadline.
On BroadNet Inference
BroadNet does not publish a separate deprecation policy or fixed notice period. Its Terms of Service (§7.4) say BroadNet “may change, suspend or retire a model or feature and, where we can, give reasonable notice on the catalogue page”, and that self-serve has no availability commitment. BroadNet’s machine-readable pricing.json gives each model a status of live or on_request. Do not rely on /v1/models for this: BroadNet describes it as catalogue metadata for external monitors, and it does not list every chat model. Our post on reading model status includes a status-checking script that you can run in CI to catch a model moving or disappearing. The provider checklist suggests asking for notice terms in writing if they matter, and keeping a second provider for critical paths. BroadNet’s Terms describe its models as open-weight models or models trained by BroadNet. For an open-weight model, the weights stay available under their licence even if BroadNet stops serving it.