Claude Opus 4.5-4.8 and Sonnet 4.6 retirement dates vs Sume ids

Anthropic lists retirement dates for Opus 4.5 through 4.8 and Sonnet 4.6. Sume's Format model field rejects those ids; the catalog row is Opus 5.5.

4 min readSume
All posts

Anthropic's deprecations page lists the Opus 4 line as active until dates between November 2026 and May 2027, with Sonnet 4.6 active until February 2027. None of those ids is a Sume catalog id, so a Format run that sends claude-opus-4-8, claude-opus-4-5 or claude-sonnet-4-6 as model gets a 400 invalid_request. The row to use on every catalog is Claude Opus 5.5.

The dates only matter for direct Anthropic calls. For a Format, the question is which ids the catalog knows, and the old four-series ids are not among them.

What are the vendor dates?

The page lists each model as active, deprecated or retired, with a not-sooner-than retirement date. Anthropic says it gives at least 60 days' notice before retiring a publicly released model.

Anthropic model deprecations page, rows read 2026-10-03
Model idStatusRetirement not sooner than
claude-opus-4-5-20251101ActiveNov 24, 2026
claude-opus-4-6ActiveFeb 5, 2027
claude-opus-4-7ActiveApr 16, 2027
claude-opus-4-8ActiveMay 28, 2027
claude-sonnet-4-6ActiveFeb 17, 2027
claude-sonnet-4-5-20250929Deprecated Sep 30, 2026Nov 30, 2026

What does Sume accept instead?

The Sume catalog code carries Claude Opus 5.5 as an always-available row, under the spellings claude-opus-5-5, claude-opus-5.5 and opus-5.5, and maps the retired claude-opus-5 onto it. Sonnet 5.5 and Fable 5.1 are catalog rows too, but they appear only where the workspace catalog lists them, so check before relying on them.

Anthropic's models overview lists Opus 5.5 at $4 input and $20 output per 1M tokens, with a 1M context window, 128K maximum output and a retirement date not sooner than September 22, 2027. Sonnet 5.5 is $2 and $10; Fable 5.1 is $10 and $50.

How does the 400 look?

The call doc says an id outside the catalog is 400 invalid_request and that the receipt echoes the id that ran. In the repository the message is model is not an Agents catalog model id, with the field named model. The request is rejected before a run starts, so no run is created for it.

Anthropic's own alias style differs from Sume's too. Sume's spellings use a dotted version for some rows and a dashed one for others, so copy the id from the catalog rather than from a vendor page.

What should I change in a saved script?

If you also call Anthropic directly, the dates in the table are the ones to schedule against, and the 60-day minimum is the floor, not the typical notice.

  • Replace any claude-opus-4-x or claude-sonnet-4-x value in a Format call with claude-opus-5-5, or remove model to use the platform default.
  • Do not rely on the vendor's retirement date as a Sume deadline; the call fails now, not then.
  • Run one test call and read the echoed model in the receipt.

What about Sonnet 4.5 and the November date?

The earliest date in the table is not an Opus one. claude-sonnet-4-5-20250929 was deprecated on September 30, 2026 and retires on November 30, 2026, with claude-sonnet-4-5's replacement listed as claude-sonnet-5-5. Among the Opus rows, claude-opus-4-5-20251101 is the first with a not-sooner-than date, November 24, 2026.

A not-sooner-than date is a floor. The page says Anthropic gives at least 60 days' notice before retiring a publicly released model, so an Active row can move to Deprecated with a firm date later. Check the page again before committing a migration calendar to a date in the table.

How do I move a call safely?

Change one saved call at a time. Send it with claude-opus-5-5, read the echoed model and the two usage fields on the receipt, and compare the cost with the old run before changing the rest. The model is more capable than the one it replaces, and a different model can use more or fewer tokens for the same input, so a cost difference is expected and worth reading rather than assuming.

Do the same for bulk runs. Each item carries its own body, so a stale model can sit in any of them; try a small bulk run first, and read the echoed model on each child receipt before queueing all 100.

Sources

Related posts

More in Models

All Models posts

Written by Sume