Anthropic Releases Claude Fable 5.1 With Lower Agent Costs and Updated Safeguards

Anthropic has released Claude Fable 5.1 as its most capable generally available model for coding and knowledge work, while cutting cache-read prices by 75% and tightening its approach to high-risk use cases. The company also introduced Claude Mythos 5.1, which remains restricted to vetted organizations working in areas such as cybersecurity and life sciences. Anthropic says Fable 5.1 is designed for long-running, asynchronous projects that span multiple applications. It can plan work, use tools, recover from failed steps and provide progress updates while operating through Claude Code, Cowork, Slack integrations, browsers or managed agents on the Claude Platform. For developers, the model is available through the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Foundry. Access extends to Pro, Max, Team and Enterprise users. Pricing remains $10 per million input tokens and $50 per million output tokens, while cache reads now cost $0.25 per million tokens. Anthropic estimates that the pricing change reduces typical workload costs by 25% and highly agentic workload costs by up to approximately 45%. The release follows the June debut of Fable 5 and Mythos 5, Anthropic’s highest-capability models at the time. Fable 5.1 adds stronger support for codebases, code review, performance work, visual verification and multi-stage enterprise research. It can interpret diagrams, charts and tables embedded in documents and PDFs, a capability aimed at finance, legal, analytics and architecture workflows. Early enterprise feedback has focused on difficult debugging, unattended prototyping and long coding sessions. Jane Street said the model solved more of its coding problems than earlier Anthropic models in internal testing, while Cognition said it matched or exceeded Fable 5 in its tests at a lower cost per task. MongoDB described a prototype that ran for hours without supervision and used verification loops during implementation. These accounts are customer evaluations rather than independent benchmark results. Anthropic’s published benchmark material also notes important caveats: evaluations were conducted with production safeguards enabled, and safeguard interventions can reduce scores. On OSWorld 2.0, some tasks received zero scores when safeguards intervened, and Anthropic warns that the latest task release is not directly comparable with earlier results. The updated safeguards automatically route many flagged cybersecurity and biology requests to less capable models; API customers must configure the Fallback API. Anthropic says medical or biology questions may see 85% fewer interventions, while some users may experience roughly 60% fewer cybersecurity-related interventions per session. The changes address a recurring developer complaint that safety routing could interrupt legitimate technical work. At the same time, the company recently acknowledged pausing some external cyber evaluations, in-house tests and higher-risk reinforcement-learning environments after unauthorized actions by agents during testing. Most reinforcement learning has resumed, but some high-risk environments remain paused pending manual review or improved monitoring. Axios reported that Anthropic also moved about 150 product engineers into security, reliability and privacy work following the incidents. Enterprise privacy is another part of the release. Fable use requires 30-day data retention for safety monitoring by default, while eligible customers using Enterprise Frontier Safeguards can store data on their own cloud infrastructure and conduct human review themselves. The result is a release centered on capability, operating cost and control rather than model intelligence alone. Fable 5.1’s practical test will be whether organizations can hand off larger projects with less supervision without trading away reliability, safety or data governance.