Claude Fable 5.1

Updated

Claude Fable 5.1 is a frontier AI model released by Anthropic on September 1, 2026, designed specifically for complex, long-horizon agentic workflows, advanced coding, and intensive knowledge work.

Key Features and Capabilities

  • Agentic Performance: The model is optimized for multi-hour, autonomous tasks that involve planning, using multiple tools, recovering from errors, and operating across various applications without constant human intervention.
  • Coding and Research: It excels at codebase-wide refactors, debugging, performance optimization, and multi-step scientific research. It is noted for being less prone to "shortcuts" and more capable of addressing root causes of problems.
  • Improved Economics: While its list pricing remains at $10 per million input tokens and $50 per million output tokens, it features a 75% reduction in cache read costs (to $0.25 per million tokens). This results in an estimated 25%–45% reduction in effective costs for typical and agentic workloads compared to its predecessor, Fable 5.
  • Technical Benchmarks: It outperforms Claude Opus 5 across Anthropic's internal benchmarks, including significant gains in agentic scientific research and complex reasoning tasks.

Availability and Variants

  • Claude Fable 5.1: The generally available version of the model, equipped with standard production safeguards. It is accessible via the Claude API, Claude web interface, Claude Code, and various cloud platforms including Amazon Bedrock, Google Cloud, and Microsoft Foundry.
  • Claude Mythos 5.1: An identical underlying model to Fable 5.1 but with more permissive safeguards designed for specialized use cases, such as cybersecurity research and the life sciences. Access to Mythos 5.1 is restricted to participants in Anthropic’s trusted access programs.

Usage Guidance

Anthropic recommends using Claude Opus 5 for most general workloads and everyday chat or drafting, while reserving Claude Fable 5.1 for demanding, long-running agentic tasks where high-level reasoning and reliability are required.