Every workflow has a plan for yes.
Reliability lives in the plan for no.
With the Fable 5 and Mythos 5 launch on 9 June 2026, Anthropic introduced safety classifiers that can end a response with a distinct refusal outcome, alongside an opt-in fallbacks parameter defining what happens when they do. Add the stop_details field from Opus 4.8, and declines went from an undifferentiated error to something you can design for.
Most teams build the happy path and treat everything else as an error to log. These releases quietly ended the era where that was the only option: a refusal is now a distinct, detectable outcome with a configurable response, which means workflows that fall over when Claude declines are a design choice, not a limitation.
A refusal is not an error any more
The classifiers introduced with Fable 5 can return a distinct refusal stop reason rather than a generic failure. That distinction matters because the two need different handling: an error wants a retry or an alert; a refusal wants a different route, a rephrase, or a human. Code that catches both in one bucket handles both badly.
Fallbacks: deciding in advance what happens next
The opt-in fallbacks parameter lets an application specify what should happen on a refusal, such as routing to a different model or flow, rather than improvising at runtime. Anthropic is candid that classifiers sometimes catch harmless requests; the Fable 5 page notes the under-5%-of-sessions figure. At that rate, a defined fallback is the difference between a non-event and a support ticket.
stop_details: knowing why it stopped
Public since Opus 4.8, stop_details returns a category and explanation for why a response ended where it did. For anyone debugging truncation or unexpected stopping, that is far more useful than inferring from the output. Reliability work runs on knowing why; this field is where why lives.
Who this matters for
Anyone running Claude unattended: agents, scheduled jobs, anything in Managed Agents or wired into a business process. Interactively, a person absorbs a refusal and rephrases. Unattended, the workflow either has a plan or it stops at 2am, and which of those happens was decided when it was built.
The design exercise, in one afternoon
List your unattended Claude workflows. For each: what currently happens on a refusal, a truncation, an unexpected stop? If the answer is “it fails and someone finds out later”, define the fallback: alternate route, safe default, or human handoff, and wire it. This is unglamorous work that upgrades a demo into a dependable system.
A workflow is production-ready when you can answer, without looking it up: what happens if Claude declines this request at 2am on a Saturday? If the answer involves a person noticing on Monday, the fallback design is not finished.
Things people
usually ask.
The classifier behaviour arrived with the Fable 5 and Mythos 5 launch on 9 June 2026, with the opt-in fallbacks parameter alongside it. The stop_details field has been public since Opus 4.8 in late May.
Rarely. Anthropic puts classifier interventions at under 5% of sessions on average, concentrated in sensitive domains like security and life sciences. The point of fallbacks is that rare times unattended still equals eventually, and eventually deserves a plan.
That is one option. A good fallback matches the workflow: route to another model, degrade to a safe default output, or hand to a human with context. The wrong fallback is an automatic retry loop that re-submits the same declined request.