Anthropic's Claude 5.1 separates capabilities from permissions, report says

Fable 5.1 is aimed at general coding, knowledge work, and agent tasks, while Mythos 5.1 is designed for vetted cybersecurity and life-science research scenarios, allowing access to deeper tool calls and more specialized workflows.

The report highlighted a test by Ramp in which Fable 5.1 ran unattended for 38 hours. During the run, the model identified a label artifact in an experiment, reprocessed the data, launched six experiment groups in parallel, and revised subsequent plans based on the returned results. Leiphone said the model changed its approach rather than mechanically following its initial plan when conditions shifted.

Leiphone's analysis argues that the release disproves the idea that a long context window alone gives an AI enough memory for sustained agent work. It says long-running tasks require explicit state management and invalidation mechanisms so outdated conclusions are not mistaken for current facts. The report points to the need for external sources of truth for code, experimental data, and task progress.

Fable and Mythos are not two different intelligence levels but two different capability profiles that can be controlled at runtime, the report added. It describes this as "capability routing": in an extended agent task, permission decisions are reassessed incrementally as the task's nature evolves, rather than only at the start.

Anthropic also demonstrated research tasks including protein design, Venus terrain reconstruction, and GPU kernel optimization. In these workflows, Claude does not replace domain-specific models; it organizes and schedules multiple specialized tools to carry out a continuous task, Leiphone said.