Anthropic launches Claude Fable 5.1 and Mythos 5.1

Anthropic has introduced Claude Fable 5.1 and Claude Mythos 5.1, two versions of the same underlying model with different safeguards and access rules. The company announced the release in September 2026. Fable 5.1 is generally available, while Mythos 5.1 is reserved for trusted-access programmes aimed at selected cybersecurity and life-sciences work.

Anthropic presents Fable 5.1 as a model for coding, knowledge work and long-running problem solving. In its published evaluations, the model scores higher than Fable 5 on several tasks, including agentic research, coding, computer use and multidisciplinary reasoning. Anthropic reports 52.6% on Terminal-Bench-Science 0.1, 55.8% on Terminal-Bench 4.0 and 73.4% on CursorBench 3.2.0. These are the company’s own results, and benchmark outcomes depend on tools, prompts, effort levels and evaluation methods.

The release also changes the economics and controls around the model. Anthropic says typical token-billed workloads will cost about 25% less than with Fable 5 because cache reads are cheaper; highly agentic workloads could be up to 45% cheaper. Its Enterprise Frontier Safeguards programme is intended to give enterprise customers zero-data-retention privacy while keeping automated misuse detection in place. The programme will roll out in phases, starting later in the autumn.

Anthropic also says Fable 5.1’s cybersecurity safeguards produce around 60% fewer interventions and allow defensive vulnerability discovery, while still redirecting penetration testing, exploit generation and some binary scanning to other models. Mythos 5.1 offers more permissive safeguards for vetted users, but remains subject to access programmes and additional restrictions.

For AI users and makers, the announcement matters because capability, price and operational boundaries are moving together. A stronger model at lower cache-read cost could make unattended coding, research and document workflows more practical. At the same time, the split between broad availability and trusted access shows that model deployment is increasingly shaped by risk controls. Companies considering Fable 5.1 should test quality and cost on their own workloads, and should treat Anthropic’s benchmark and safety claims as claims to assess rather than guarantees.