Zero Data Retention (ZDR) is an arrangement some enterprises negotiate with Anthropic under which API inputs and outputs are not retained after the request completes. It's a popular ask from security and legal teams. But it interacts badly with one specific model: Claude Fable 5 requires at least 30 days of data retention and is not ZDR-eligible. If your organization is configured for ZDR — or any retention window below 30 days — every request to Fable 5 returns a 400 invalid_request_error, no matter what the payload contains.
Why this 400 is so confusing
A 400 is documented as "there was an issue with the format or content of your request," and it is not retryable — the standard advice is to fix the request and resend. That advice sends engineers down exactly the wrong path here. They will diff payloads against working Sonnet 5 calls, strip parameters one by one, and reduce the request to a bare minimum "hello" — and still get a 400, because nothing in the request body is being evaluated. The gate is organizational: the model requires retention your org's configuration forbids, so the API rejects the call before any content validation matters.
The telltale signature is the pattern, not the message: every Fable 5 request from the org fails with 400, while the same requests succeed instantly when you change only the model ID to claude-opus-4-8 or claude-sonnet-5. When you see that pattern, stop debugging the payload and check your organization's data-retention configuration.
Why the requirement exists
Anthropic's documentation ties the requirement to Fable 5's safety posture: it is Anthropic's most capable widely released model, and its documentation notes additional safeguards — Microsoft's Foundry docs, for example, describe extra input/output safety classifiers applied to Fable 5. Retention gives Anthropic's trust-and-safety processes something to review. Notably, the docs distinguish the feature from the model: adaptive thinking as a capability is ZDR-eligible, but Fable 5 and Mythos 5 themselves are not available under ZDR. The restriction is about the model, not about any parameter you send.
What your options are
Option 1: use a different model. If ZDR is contractually non-negotiable for your organization, Fable 5 is off the table. Claude Opus 4.8 is Anthropic's recommended starting point for complex enterprise and agentic work and doesn't carry the 30-day retention requirement.
Option 2: adjust retention for the org (or a separate org). If Fable 5's capabilities justify it, the change is a governance decision: move the organization to at least 30-day retention, or run Fable 5 workloads from a separate organization with standard retention while keeping ZDR-sensitive workloads where they are. This is a conversation with your Anthropic account team and your own security/legal stakeholders — no API change is involved either way.
Option 3: check the platform angle. On Claude Platform on AWS, the data-retention policy matches the first-party Claude API, and ZDR is available on request through an Anthropic account representative — so the same Fable 5 constraint logic applies there. Organizations with strict retention requirements should map which workloads need which models before committing to a retention posture, because retro-fitting is an organizational change, not a code change.
A checklist for the on-call engineer
When Fable 5 returns 400s: (1) confirm the identical request succeeds on another current model; (2) confirm the failures are universal for Fable 5 across your org's API keys, not specific to one payload shape; (3) read the error body's message field — log it in full, since it identifies the violation; (4) escalate to whoever owns your Anthropic org settings with the finding "our retention configuration is below Fable 5's 30-day requirement." That one sentence saves the next team a day of payload archaeology.
Where to go next
For the request-shaped 400s that are about your payload, see model-specific 400 errors. For the broader error landscape, start with the API error code reference.