Claude Fable 5 (claude-fable-5, $10 input / $50 output per million tokens) became generally available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry on June 9, 2026. Most of its migration notes are the usual model-upgrade fare — adaptive thinking always on, no assistant prefill, safety-classifier refusal categories. One is different in kind: Fable 5 requires a minimum 30-day data retention window and is not available under Zero Data Retention (ZDR). Unlike a missing feature you can code around, this is a policy constraint that follows the model onto every platform.
What ZDR is, and how the constraint bites
Zero Data Retention is an arrangement some organizations negotiate with Anthropic under which API inputs and outputs are not retained after the request completes — a common ask from legal, healthcare, and financial-services teams. The friction with Fable 5 is absolute: on an organization set to ZDR, or any retention period below 30 days, every Fable 5 request fails with a 400 invalid_request_error, regardless of what is in the payload.
Anthropic's documented paths forward for ZDR organizations are to contact their account team or to configure retention per workspace — carving out a workspace with 30-day retention for Fable 5 traffic while the rest of the organization stays on ZDR. Whether a per-workspace carve-out is acceptable is a compliance question, not an engineering one; workloads whose data genuinely cannot be retained simply cannot use Fable 5.
How this plays out per platform
Fable 5 is on all five surfaces — claude-fable-5 on the first-party API, Claude Platform on AWS, Vertex AI, and Foundry, and anthropic.claude-fable-5 on Bedrock — but the retention question looks different depending on who operates the surface.
Claude API (1P) and Claude Platform on AWS. These share the same data retention policy, and ZDR is available on request through an Anthropic account representative — which means the Fable 5 conflict applies directly on both. The per-workspace retention approach is the documented mitigation. On Claude Platform on AWS, note that AWS IAM gives you an enforcement tool the API itself doesn't: policies can deny specific actions on ZDR-sensitive workspaces so retained-data features can't be used there by accident.
Bedrock, Vertex AI, Foundry. On the cloud-operated platforms, data handling is governed by the cloud provider's own terms rather than an Anthropic ZDR agreement — Bedrock runs on AWS-managed infrastructure with zero Anthropic operator access, Vertex data handling follows Google Cloud's data governance, and on Foundry Anthropic acts as an independent data processor under its DPA (with Azure-hosted deployments keeping prompts and completions within Azure). If your regulatory posture depends on specific retention behavior for Fable 5 on these platforms, confirm it with your cloud provider and account team; the platforms inherit their providers' compliance posture, and the public documentation does not spell out a Fable-5-specific retention mechanism for them. One practical Foundry note: claude-fable-5 ships as a Hosted-on-Anthropic Preview there, and pay-as-you-go subscriptions default to a quota of 0 — so access requires quota arrangements before retention even becomes the question.
Designing around the constraint
Route by data classification, not by model preference. The clean pattern is a policy layer that assigns models by workload sensitivity: ZDR-bound workloads run Claude Opus 4.8 ($5/$25, Anthropic's recommended starting point for complex enterprise and agentic work, and fully available without the 30-day requirement), while retainable workloads may opt into Fable 5 where its capability is worth 2x the token price.
Audit the rest of your pipeline while you're at it. ZDR-sensitive workloads face other non-eligible features regardless of model: the Message Batches API and the Files API are both not ZDR-eligible, while token counting, PDF processing, and citations are. A workload that already avoids batches and file uploads for ZDR reasons has less to untangle here.
Keep the fallback boring. Fable 5 and Opus 4.8 share a tokenizer, so token counts and costs stay roughly comparable when a request is rerouted between them — one less variable when your router downgrades a request from Fable 5 to Opus 4.8 because of its data classification.
Where to go next
For the broader retention landscape, see retention and deletion for regulated workloads. The platform overview summarizes who operates each surface — the fact that decides whose retention terms apply.