Claude Fable 5 is a large-scale language model developed by Anthropic, representing the first general release of the Mythos-class architecture. Launched in June 2026, the model is designed for high-autonomy agentic work and complex reasoning tasks that exceed the capabilities of the previous Opus-tier systems. It is positioned as a frontier model for long-horizon knowledge work, scientific research, and advanced software engineering.
The model features an adaptive reasoning system controlled via a dedicated effort parameter. This allows for the modulation of thinking depth across discrete levels, including "low," "medium," "high," and "max." At the "max" effort setting, the model prioritizes exhaustive reasoning and exploration, which has been shown to improve performance significantly on difficult benchmarks such as FrontierCode Diamond and AA-Omniscience, where it demonstrates the ability to convert increased token budgets into higher task accuracy.
A distinctive characteristic of the model is its integrated safety fallback mechanism. To balance frontier capability with safety, Claude Fable 5 employs conservative classifiers that monitor for potentially harmful queries in domains like cybersecurity and biology. When a request is flagged, the system automatically routes the task to Claude Opus 4.8. This fallback is intended to be seamless, with Anthropic reporting that it occurs in fewer than 10% of sessions while maintaining session continuity.
Claude Fable 5 supports a 1 million token context window and a maximum output limit of 128,000 tokens. It is a multimodal system capable of processing text, images, and PDF documents as inputs. The model is optimized for autonomous tool use and is capable of navigating complex user interfaces and executing multi-step agentic workflows with minimal human intervention.