Safety Routing Fallback
A safety classifier can silently reroute a "sensitive-looking" request to a lower-tier backup model. The source describes a mechanism (attributed to Anthropic's Prompting Claude Fable 5 / Refusals and fallback docs) in which Fable 5 runs a quick safety check before answering; if a request trips it — anything resembling offensive cybersecurity, dangerous biology/life-sciences, or asking the model to reveal/narrate its own private ("summarized") reasoning — the task is handed off to a less-capable backup (Opus 4.8) rather than refused outright. The operationally important consequences:
- It can be silent. In the Claude app and Claude Code there is no on-screen notice that a different model answered; on the raw API the response is marked. So you can get a weaker answer without knowing it switched.
- Benign work can trip it. Ordinary coding, debugging, and security questions "can occasionally get caught too" — the classifier is not a precise oracle.
- The Fable-specific prompting rule that follows: a standing "explain your reasoning step by step" line — especially saved in a system prompt or skill — is exactly a reasoning-extraction pattern and can trigger the reroute. Prefer "just give the answer and think privately; if you need the reasoning, read the thinking output" over making the model narrate its reasoning.
The framing the source lands on: a handoff is not a failure or a surprise bill — you get a capable Opus 4.8 answer at roughly half Fable's price (a declined request that writes nothing is free), losing only Fable's edge on long/hard tasks. This is the only Fable-specific habit among the six; the other five are general prompting.
Claims
- A safety/capability classifier that reroutes rather than refuses can silently downgrade the model answering your request, so the operator must know the reroute exists to reason about their results. observation — a mechanism claim about routing behavior; its importance is that a silent capability change is invisible in the app but real in the output.
- Don't make the model narrate its private reasoning (a standing "explain your reasoning" line, especially in a system prompt/skill) on a model with a reasoning-extraction safeguard — ask for the answer and read the thinking output instead. best practice — context: Fable-specific; this is contingent on that model's safeguards and may not apply to models without reasoning-extraction classifiers, so it must not be generalized into a universal rule.
- Keep genuinely sensitive work (cyber/biology) on the intended lower-tier model by default, and word security questions plainly, to avoid accidental reroutes. best practice — context: operators whose legitimate work brushes the classifier's domains; configure server-/client-side fallback deliberately rather than being surprised by it.
- A reroute costs less, not more — the backup is cheaper and a pre-write refusal is free — so the only thing lost is the top model's capability on hard tasks. observation — an attributed pricing/behavior claim (Opus ≈ half of Fable); groundable and flagged for verification.
Related
- Reasoning Effort Control — a cheaper backup answering a routed request is the same cost/capability tradeoff seen from the routing side.
- Concise Prompting — "stop asking it to show its reasoning" is both a routing-avoidance rule and an instance of cutting instructions a capable model doesn't need.
- Agent Supervision — a silent model swap is a legibility gap: the operator can't supervise a capability change they can't see (raw-API marking is the legible channel).
- Intent Context — the classifier judges the apparent intent of a request; clear, plainly-worded intent reduces false trips.
- Distillate: How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch