Particle.news
Download on the App Store

Anthropic Opens Public Access to Mythos-Class AI With Guardrailed Fable 5

Anthropic widens access to its most capable model by routing risky cybersecurity, biology and chemistry prompts to Opus 4.8; full Mythos 5 remains limited to vetted partners.

Overview

  • Anthropic announced the rollout Tuesday, making Claude Fable 5 generally available while offering a less-restricted Claude Mythos 5 only to vetted Project Glasswing partners and select researchers.
  • Fable 5 runs new classifier-based guardrails that detect sensitive queries and automatically route them to the less-capable Claude Opus 4.8, a fallback that Anthropic says triggers in under 5% of sessions on average.
  • The company says Fable 5 shows big gains on long, complex coding, vision, reasoning, and memory tasks and cites third-party benchmarks that rank it above prior public models.
  • Anthropic reported more than 1,000 hours of internal and external red‑teaming with no universal jailbreaks, set a 30-day traffic retention policy for safety monitoring, and priced Fable/Mythos at $10 per million input tokens and $50 per million output tokens.
  • The move follows months of Project Glasswing testing that uncovered thousands of high‑severity vulnerabilities and exposed a patching bottleneck, and it raises questions about scaling defenses, oversight and how trusted‑access programs will expand.