Claude Fable 5 and Claude Mythos 5 Put AI Governance in the Spotlight

Anthropic has released Claude Fable 5, a new AI model that the company says delivers its strongest performance yet across software engineering, reasoning, vision, scientific research, and long-duration tasks.

Alongside the launch, Anthropic introduced Claude Mythos 5, a version of the same underlying model that is currently limited to approved cybersecurity partners and select research organizations.

The announcement goes well beyond benchmark performance. Anthropic also details a new framework for deploying advanced AI systems, including safeguards, access restrictions, monitoring requirements, and trusted-access programs for capabilities the company considers higher risk.

Why It Matters: The launch offers insight into how advanced AI systems may be governed once they begin handling work with security, research, and operational implications. Anthropic is pairing new model capabilities with safeguards, monitoring requirements, retention policies, and access controls, illustrating how governance is becoming a larger part of advanced AI deployments.

• Fable 5 and Mythos 5 Share the Same Foundation Model: Fable 5 and Mythos 5 are built on the same underlying model, though they are being released under different access conditions. Anthropic describes Fable 5 as its most capable publicly available model, while Mythos 5 remains limited to approved cybersecurity and research organizations. The distinction is tied largely to access and safeguards, giving Anthropic a way to make advanced capabilities available to a wider audience while retaining tighter controls around certain use cases.

• A Focus on Long-Running Tasks and Knowledge Work: Much of the announcement centers on work that unfolds over hours, days, or even longer periods. Anthropic points to software engineering projects, financial analysis, document review, and vision-based tasks that require maintaining context across large amounts of information. The company also emphasizes improvements in memory and persistence, describing scenarios where the model used stored notes and prior work to improve results over time.

• Biology and Drug Discovery Feature Prominently: Scientific research occupies a significant portion of the release. Anthropic describes work involving protein design, biological hypothesis generation, and genomics research conducted with limited human input. Several examples focus on tasks that would traditionally require specialist expertise, though the company notes that some of these same capabilities raise questions about how advanced scientific systems should be deployed and monitored.

• New Controls Govern High-Risk Requests: Those concerns help explain why safeguards feature so prominently in the launch. Fable 5 includes classifiers that evaluate requests related to cybersecurity, biology, chemistry, and model distillation. When those systems are triggered, responses are routed to Claude Opus 4.8 instead of Fable 5. Anthropic says the safeguards activate in fewer than 5% of sessions and describes ongoing efforts to reduce false positives while maintaining protections against misuse and jailbreak attempts.

• Anthropic Introduces Tiered Access: The rollout also introduces a more structured approach to access. Organizations participating in Project Glasswing can use Mythos 5 with certain cybersecurity restrictions removed, while similar programs are planned for biomedical researchers. Anthropic is also implementing a 30-day retention requirement for Mythos-class traffic to support safety investigations and monitoring. Together, these measures suggest that access to advanced AI capabilities may increasingly be governed through permissions, oversight requirements, and trusted-access programs instead of broad availability alone.

Go Deeper -> Claude Fable 5 and Claude Mythos 5 – Anthropic