Anthropic Split One Model Into Public Fable and Restricted Mythos

“They trained one model, then sold the censored version cheaper and locked the real one behind a trusted-access gate.”

The Story

Anthropic launched Claude Fable 5.1 (generally available) and Claude Mythos 5.1 (same underlying model, restricted trusted-access for cybersecurity and life sciences). Fable is cheaper—roughly 25% less on typical work, up to 45% on long agentic tasks—with improved safeguards and enterprise privacy options. Early testers report it staying coherent on multi-day unattended coding and research.

Why It Matters

One intelligence, two doors: the “fable” version for everyone and the “mythos” version with dangerous capabilities left on. Builders get cheaper overnight agents; the strange part is policy turning the same weights into public vs. restricted personalities.

Evidence

Anthropic official announcement (Sept 1, 2026)

Sources

Daily scan: 2026-09-02