Fable models are unavailable to zero-data-retention organisations

The error

An organisation configured for zero data retention sends a well-formed request to a Fable model and gets a 400 invalid_request_error instead of a response — not a rate limit, not a malformed request, but a refusal tied to the account's retention setting:

In order to access this model, your organization or workspace must have data retention enabled.

Which models this applies to

Fable 5 and Fable 5.1 — and their Project Glasswing counterparts, Mythos 5 and Mythos 5.1 — are what Anthropic calls Covered Models: they require the standard data-retention window, and are not available under zero data retention unless Anthropic has expressly authorised it. When this page was first written, Fable 5 was the only model in that position; its successor inherited the requirement unchanged. The constraint is a property of the models, so there is no per-request setting that relaxes it.

A 404 is worth checking the same way

An early draft of Anthropic's launch documentation described this case as a 404 with the model hidden from the models list; the final behaviour is the 400 above. If you see a 404 on a Fable id, check the organisation's retention configuration before debugging the payload.

What to route to instead

For a zero-retention workload, the documented option is Opus 5, which is available under zero data retention. For Opus 5.5 the migration guide records nothing new on retention and says to treat it as Opus 5 — check with your account team before a zero-retention workload moves onto it. An organisation that needs a Fable model can ask its account team about the authorised path, or enable retention for a single workspace that serves only that workload.

Where else this surfaces

Check any model-picking logic — a router, a fallback chain, an eval harness cycling through models — for a hardcoded Fable id. That is the one place a zero-retention account meets this error even when nobody chose Fable for the task. See zero data retention and which models support it for building that filter into model selection.