Claude Mythos 5
anthropic‘s unrestricted frontier model (9 June 2026), the sibling of claude-fable-5 — per the official announcement, the identical underlying model with safeguards lifted in some areas, for authorized users. The “Mythos-class” capability ceiling; Fable 5 is the same model wrapped in safety classifiers for general use.
What distinguishes it
- Same model, fewer guardrails. Where claude-fable-5 runs three safety classifiers (cybersecurity, biology/chemistry, distillation) that fall back to claude-opus-4-8, Mythos 5 removes that friction for verified professionals who need unimpeded access.
- Restricted distribution (the opposite of a public launch). Initial access via Project Glasswing (US-government collaboration); cybersecurity partners can upgrade from Mythos Preview immediately; a biology-focused trusted-access program is planned. Broader access TBD.
- Pricing: $10/$50 per Mtok (same as Fable 5).
- Capability evidence: matched or exceeded skilled human operators on protein-design tasks (the headline Mythos-only result, and exactly the dual-use domain Fable gates).
- Alignment: automated assessment found misaligned behavior “low, similar to Opus 4.8.”
Role in this wiki — a new shape for the safety story
Mythos 5 is the conceptual key to the substrate’s safety-tier split: Anthropic now ships capability and safety as separate axes — one model, two access tiers gated by who you are rather than what the model can do. This is a notable shift from the claude-opus-4-8 story (where safety = honesty, a property of the single released model): now safety is deployment policy (classifiers + trusted-access gating + 30-day retention), not just model disposition. It sharpens the synthesis open question about an auto-maintaining wiki’s substrate: the wiki runs on the guardrailed tier (claude-fable-5), and the existence of an ungated sibling makes “whose model, under what policy” an explicit variable.
Related
claude-fable-5 · claude-fable-5-mythos-5-announcement · claude-opus-4-8 · anthropic · synthesis