Breaking

Anthropic Unveils Public Fable 5 and Restricted Mythos 5 From One Model

June 12, 2026 at 11:17 EDT

Foundation Models
AI Agents
Security

On June 9, 2026, Anthropic announced two models built on its top-tier "Mythos-class" foundation: the publicly available Claude Fable 5 and a less-restricted twin, Claude Mythos 5, with some safety guardrails removed. The two share the same underlying model and differ only in their safety measures, with Mythos 5 limited to vetted cyber defense teams.

June 9 · Anthropic

One Model, Two Doors: Frontier AI Becomes a Gated Utility

Anthropic split its new Mythos-class foundation into two: Claude Fable 5, public with hardened safeguards, and Claude Mythos 5, restricted to vetted cyber teams with guardrails removed. Same base — different access to full power.

90%+

First score above 90% on the core long-horizon analytics benchmark (+10 pts over Opus)

Token context length, shared across both configurations

$10/$50

Input / output per million tokens — far cheaper than Mythos Preview

SWE-Bench Pro — software engineering

Higher is better · score in %

80.3%

Fable / Mythos 5

69.2%

Opus 4.8

58.6%

GPT-5.5

Fable 5 · Public

Generally available on claude.ai (paid), the API, and Amazon Bedrock — Mythos-class capability with built-in safeguards for general users.

Mythos 5 · Restricted

Cyber guardrails removed. Offered only to vetted teams in Project Glasswing — later expanding to biology researchers.

High-risk fallback mechanism

High-risk query
cyber · bio · chem · health

→

Flagged & refused
API refusal notice

→

Auto-fallback to Opus 4.8
lower performance

Developers say

A clear step up on long-horizon, complex agentic work — strong on large-scale code migrations and multi-day projects.

Critics warn

Guardrails feel too strict, flagged queries degrade to Opus 4.8 — and gating full power signals AI as a controlled utility.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.