Retirement notice: Anthropic 60 days, OpenAI 6 months. Pin the model?

Both vendors publish a floor. Anthropic says it gives at least 60 days' notice before retiring a publicly released model, and OpenAI says at least six months for a generally available model and at least three months for a specialized variant. For a saved Sume Format, the better question is who moves the id when it retires, and the answer is the Sume catalog, not the vendor.

So pinning a model in a saved call buys you a fixed choice until the catalog changes it, and omitting it buys you whatever the platform default is at the time. Which one you want depends on whether you value stability or tracking.

What notice does each vendor promise?

The figures below are the policy statements on each vendor's deprecations page. They are minimums, not what a given retirement used.

What happens to a retired id on Sume?

The catalog code in the Sume repository keeps retired spellings and points each at a successor: the retired Opus 5 resolves to Opus 5.5, Sonnet 5 to Sonnet 5.5, Fable 5 to Fable 5.1, and gpt-5.6-sol to gpt-6-sol. The call doc names the last one: a request for gpt-5.6-sol runs on gpt-6-sol, and the receipt echoes the id that ran.

That is not universal. An id that was never a catalog id, such as an older OpenAI or Claude model, returns a 400 rather than a successor. The alias list is a decision recorded in the catalog for specific retirements.

Pin or omit?

Omit model when the Format does not care which LLM orchestrates it. The platform default applies, and you do not carry an id that can go stale. Pin it when you compared two models on your own input and chose one, because a comparison is the only reason to prefer a fixed id; the receipt's echoed model then tells you if something moved underneath you.

The alias route has a cost: a pinned retired id keeps working only while the catalog keeps its alias, and you learn about a change from the receipt, not from a notice.

A rule for a team

The runs doc shows where the echoed model appears on a receipt.

• Keep a short list of every model value your saved calls and bulk items use.
• Check the list against the two deprecation pages monthly; the pages carry the dates.
• After any catalog or vendor change, run one test call and compare the echoed model with what you expect.
• Remove the field from calls that never needed it.

What do the real dates look like against the floors?

The deprecation pages show how the floors play out. OpenAI's 2026-10-01 entry lists gpt-5.1, gpt-5.3-codex and gpt-5.4-nano for an April 1, 2027 shutdown, six months later. Anthropic's rows carry retirement dates well beyond 60 days: the Opus 5.5 row, for example, is listed as not retiring sooner than September 22, 2027.

Long notice is not a reason to leave an old id in place. It is a window in which to compare, test and migrate on your own schedule, which is easier when each saved call is on a list you keep.

What about a Format that must not change?

Some Formats are reviewed by a person and approved as a unit. For those, a pinned id plus a receipt check is a reasonable control: if the echoed model ever differs from the pinned one, the run is flagged for review before its output ships. Where nobody reviews the choice of LLM, a pinned id is only a future 400 waiting for a retirement, and omitting the field is simpler.