What happened
On June 9, 2026, Anthropic released Claude Fable 5 for general use and Claude Mythos 5 for a restricted group of cyberdefenders, alongside a detailed system card. Key findings: Mythos 5 is the most capable cyber model evaluated to date, scoring far ahead of Opus 4.8 on exploit development; Fable 5 uses cybersecurity classifiers that route ~5% of sessions to Opus 4.8 when cyber-related queries are detected; Mythos 5 achieved the lowest (best) result on an external Gray Swan prompt-injection benchmark; breaking the cybersecurity safeguards is assessed as 'extremely difficult though not impossible'; and unsafeguarded Mythos 5 can 'significantly uplift well-resourced threat actors' on chemical and biological research. Fable 5 is priced at $10/1M input / $50/1M output tokens — less than half the price of Mythos Preview.
Why it matters
This is the first Tier 2 system card to quantify dual-use cyber risk at Mythos scale and document the specific safeguard architecture used to manage it. The public release of Fable 5 at Mythos-class capability levels — even with classifiers — means the threat model for AI-assisted cyberattacks has materially shifted: defenders must now assume adversaries have or will have access to models of comparable capability. The Gray Swan prompt-injection benchmark score is a useful external reference point for agentic safety evaluation frameworks.
Action needed
Read the Fable 5/Mythos 5 system card to update your organisation's AI threat model; specifically, use the Gray Swan benchmark and METR evaluation results as reference points for your own red-teaming scope, and brief clients on the implication of Mythos-class models now being accessible to the general public via Fable 5.