Anthropic has introduced Claude Fable 5.1 alongside a specialized variant called Claude Mythos 5.1 for frontier research and programming. The new flagship architecture brings substantial price cuts to prompt caching while outperforming older versions in autonomous software troubleshooting and biology. While Fable 5.1 is available immediately for all developers, access to Mythos 5.1 remains restricted to vetted security and life sciences teams.
Running large software agents gets expensive fast. Anthropic addressed this problem by slashing prompt cache reads by 75%, bringing the rate down to $0.25 per million tokens. Standard token rates stay unchanged at $10 per million for inputs and $50 per million for outputs. The overall impact drops operational bills by roughly 25% for standard tasks, while deep agent loops can see savings climb as high as 45%.
Under heavy development tasks, the model avoids shortcuts and tracks down root causes across multiple software repositories. During evaluation by quantitative teams at Millennium, the system solved an elusive crash that human engineers spent years attempting to diagnose.
Every model I tried missed it. Claude Fable 5.1 was the first to find it. It disassembled an external vendor library, matched it against the core dump, and traced the crash to a bug in that library.
Benchmark scores reflect these qualitative gains. The model reached 55.8% on Terminal Bench 4.0, while the specialized Mythos 5.1 build reached 60.9%. On real world developer workflows tested through CursorBench, the system scored 73.4%, demonstrating strong ability to verify its own work without human supervision.
The architecture also showed strong gains in academic settings. In laboratory evaluations of protein design, Mythos 5.1 achieved a success rate near 50% across 12 targets, creating high affinity binders that far exceeded standard benchmark competitions. When applied to machine learning infrastructure, the system wrote custom GPU kernels that accelerated open source genomic models by up to 2.5 times on NVIDIA H100 hardware, cutting computational expenses by up to 60%.
Enterprise data protection gets an update through a framework called Enterprise Frontier Safeguards. This allows companies to keep data within their own private cloud servers rather than Anthropic infrastructure, providing zero data retention guarantees. Cybersecurity safeguards now produce 60% fewer false alarms, allowing security professionals to locate vulnerabilities without triggering automated blocks. The release also includes an invisible watermark designed to comply with European transparency rules without altering output quality.
