Fable 5 and the Model That Reroutes Itself
On June 9, Anthropic put a Mythos-class model in front of the public. Claude Fable 5 shipped with a 1M token context window, up to 128k output tokens, and state-of-the-art scores on nearly every benchmark they tested. Most of the coverage is about the scores. I care about something quieter in the release notes.
Fable 5 ships with safety classifiers that can decline a request. Its sibling Mythos 5, available only to approved customers, does not. And when Fable 5 detects a query in a high-risk domain, it doesn't just refuse. It reroutes the request to Claude Opus 4.8.
That is a model deciding, mid-request, which model should answer.
Why this matters for enterprise
For a year, "model governance" in the enterprise meant a config file. You picked a model, set a temperature, and wrote a policy doc nobody read. The control lived outside the model.
Rerouting moves part of that control inside the model. The classifier is now a gate, and the gate has a fallback path instead of a dead end. From a controls standpoint that is a much better primitive than a hard refusal, because a refusal in production is an outage and a reroute is a degraded-but-working response.
The finance version of this is obvious once you've lived through a SOX audit. You do not want a model that silently does the risky thing, and you do not want one that hard-stops and breaks the workflow. You want a documented downgrade path: when the request crosses a risk line, the system quietly drops to a more constrained behavior and logs that it did.
The takeaway
The headline is a smarter model. The durable change is that the safety decision and the fallback are now part of the inference path, not a wrapper around it.
If you're building on top of this, stop treating the model as a single endpoint. Treat it as a router with policy baked in, and make sure your own logging captures which path a request took. The day an auditor asks "why did this request get a different answer," the reroute log is the paper trail.
Pricing settled at $10 per million input tokens and $50 per million output on June 23. The capability is worth it. The architecture is worth more.