Series start. On June 9, 2026, Anthropic launched a Mythos-class stack in two forms: Claude Fable 5 for general use, and Claude Mythos 5 for a narrow trusted-access path. Anthropic says they share the same underlying model; the difference is safeguards and who can call which configuration.
This is the event piece in Brocker’s Anthropic Frontier series. Later pieces cover Glasswing, the June access pause, Opus 5, and the consequence file — not rewrites of this announcement. Summer competitive tempo: frontier cadence readout.
What Anthropic shipped
- Claude Fable 5 — Mythos-class model made safe for general use, with classifiers that can block or fall back on high-risk domains such as cybersecurity and biology (Anthropic’s launch materials route many blocked requests to Claude Opus 4.8; later Opus 5 materials update biology fallback routing toward Opus 5).
- Claude Mythos 5 — same underlying model with some safeguards lifted; initially through Project Glasswing for cyber defenders and infrastructure partners; not generally available.
- Pricing (launch) — Anthropic listed Fable/Mythos 5 at $10 / $50 per million input/output tokens — less than half Mythos Preview.
Anthropic positions Fable 5 as state-of-the-art on many capability benchmarks it tested, with particular emphasis on software engineering and long, complex tasks — while warning that releasing that capability without safeguards creates serious misuse risk.
Why the split matters
Most labs ship one public flagship. Anthropic shipped a capability twin: public track with blockers, partner track with lifted classifiers. That design choice is the spine of this series — unpacked in the consequence file.
Our take
In our view, Fable/Mythos 5 is less a branding exercise than an admission that frontier coding/cyber capability now forces product architecture: who gets the raw model, and who gets the safeguarded twin.
Series: Anthropic Frontier · next: Project Glasswing.
