Claude Fable 5.1 updates long-running agents, science and context economics

[{"h":"CONFIRMED","p":"Anthropic updated Claude Fable 5.1 and Claude Mythos 5.1 to make long-running agents more viable as a product, not merely to post a higher benchmark score. Fable 5.1 is available to all customers through the API and major clouds. Mythos 5.1 uses the same base model but is restricted to trusted-access programs for cybersecurity and life sciences. The release combines higher capability, much cheaper cache reads and a migration that needs care."},{"h":"The discount that reaches the bill","p":"Input costs $10 per million tokens and output costs $50. The cut is in prompt-cache reads, down 75% to $0.25 per million. Anthropic estimates savings of about 25% on typical workloads and up to 45% on heavy agentic work. That favors agents rereading instructions, documents and history across many calls. It is not a cheap model: fresh tokens still cost twice Claude Opus 5 on input and four times as much on output."},{"h":"Frontier capability, stricter integration","p":"The model has a 1 million token context window, up to 128,000 output tokens and adaptive thinking always enabled. Its API ID is claude-fable-5-1, available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry. The documentation recommends starting with Opus 5 and reserving Fable 5.1 for tasks where high-effort evaluations still fall short. Migration breaks forced tool use with any or tool, which now returns a 400 error. Earlier models cannot read thinking blocks produced by Fable 5.1, and editing earlier turns can invalidate them."},{"h":"The part that actually changes agents","p":"The update adds per-message effort, turn-scoped system messages, readable updates between tool calls and statistical text watermarking. The most useful detail is the ability to raise effort for a hard step and lower it later without invalidating the cache. The trickier detail is greater variability in parallel tool calls: loops that once performed several independent reads may spend more turns unless the prompt explicitly requests batching."},{"h":"Science and safeguards in one package","p":"Anthropic reports scientific research advances including a high-resolution map of one third of Venus and, in Mythos, protein binders with a hit rate near 50% across 12 targets. These are company results, not independent proof of autonomous discovery. On safeguards, Anthropic says cyber false-positive interventions fell 60% and defensive vulnerability discovery is now allowed, while exploit creation remains blocked. Enterprise Frontier Safeguards, which promises customer-controlled data storage, begins phased rollout in the fall."},{"h":"The MaxAssistant reading","p":"This release is material, but the most concrete gain for small businesses is not the ranking. It is context economics and the ability to sustain long work. Test it now on agents that repeat substantial history, with tool choice set to auto, a token cap, failure cases and real transcripts. Do not move production because of the word frontier. Do not assume zero data retention either: the documentation says 30-day retention unless Anthropic expressly authorizes an exception."},{"h":"Sources","p":"Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1: https://www.anthropic.com/claude-fable-and-mythos-5-1 | Claude Platform, Claude Fable 5.1 overview: https://platform.claude.com/docs/en/models/fable-5-1/overview | Claude Platform, What is new in Claude Fable 5.1: https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1 | Claude Platform, Migration guide: https://platform.claude.com/docs/en/models/fable-5-1/migration-guide"}]