Anthropic ships fallbacks parameter for Claude Fable 5, automatically rerouting safety-refused requests to alternative models

Alongside the June 9, 2026, launch of Claude Fable 5 and Claude Mythos 5, Anthropic shipped a new beta API parameter called fallbacks that lets developers automatically reroute refused requests to a different model — rather than surfacing a refusal error to end users. The feature addresses a behavioral shift introduced by Fable 5's built-in safety classifiers, which can decline requests during generation in ways that earlier Claude models did not.

What's new

Claude Fable 5 is the first Claude model to run safety classifiers on requests and during response generation as a core architectural feature. When a classifier blocks a request before any output is generated, the Messages API returns stop_reason: "refusal" and charges nothing for that call.

The fallbacks parameter extends this behavior: when included in a Fable 5 request and a refusal fires, the API automatically re-runs the request on a specified fallback model at that model's standard rates. The parameter is "in beta on the Claude API and Claude Platform on AWS" but is "not supported on the Message Batches API." Assistant prefill and manual extended thinking budgets — neither of which is supported on Fable 5 — trigger 400 errors rather than the fallbacks path; fallbacks applies specifically to safety-classifier-triggered refusals.

Fable 5 refusals return a stop_details.category field identifying the reason: "reasoning_extraction" (blocked for attempts to reverse-engineer model outputs), "cyber", "bio", or null. The fallbacks logic fires on any of these classifier-triggered paths.

Context

Earlier Claude models could refuse requests, but refusals were not embedded in the model-serving infrastructure as a distinct, billable event with its own stop_reason. The shift with Fable 5 reflects a design choice: the safety classifiers run at the model-serving layer — not just as post-processing — meaning refusals can occur during generation, not only at the start.

The June 2 Anthropic API update had already established that refusal-without-output calls are free across all models. The fallbacks parameter goes further by giving developers a declarative fallback path rather than requiring application-side retry logic.

For Claude Managed Agents, Anthropic separately added scheduled deployments and environment variable credential support in the same June 9 release. The fallbacks parameter is standard Messages API functionality and does not require the Managed Agents beta header.

Why it matters

The fallbacks parameter solves a friction point that Fable 5's new safety architecture introduced: a refused request no longer needs to be caught and re-routed in application code. Developers building on Fable 5 who want high-quality output for most requests — but need a fallback path when the classifier declines — can now declare that fallback inline in the API call.

This matters most for high-volume production applications where a bare stop_reason: "refusal" response would produce a visible gap in the user experience. With fallbacks, the developer specifies a recovery model (such as Claude Opus 4.8 or Claude Sonnet 4.6) and the API handles the re-run transparently.

It also signals how Anthropic is thinking about the safety-capability tradeoff at the API layer: rather than forcing a binary choice between Fable 5's capabilities and its refusal rate, the fallbacks mechanism lets developers tier their models and handle edge cases at the infrastructure level rather than in userland retry logic.

The parameter is in beta on the Claude API and Claude Platform on AWS. It is not supported on the Message Batches API.