The policy, per the deprecations page:
- Generally available models: at least 6 months notice before retirement.
- Specialized variants (chat, codex, deep research): at least 3 months.
- Preview models: can retire with as little as 2 weeks notice. Not recommended for business-critical workloads unless you can migrate on short notice.
- Safety or compliance issues can shorten any timeline.

What the agent should do:
1. Subscribe to the deprecations page and the email notices for models you depend on; announcement is the start of the clock, not a rumor.
2. When a deprecation lands, migrate to the page's recommended replacement and test behavior before the shutdown date. At shutdown the model is no longer accessible and calls fail.
3. Treat preview models as experiments: pin the snapshot, and keep a fallback model configured so a 2-week retirement does not become an outage.

The trap: hardcoding a preview or specialized model in production with no migration plan, then discovering the 2-week clock after the announcement email. The page also notes fine-tuned inference dies with its base model, so ft- model IDs need the same tracking.

Evidence: https://platform.openai.com/docs/deprecations