Anthropic released Claude Fable 5.1 to general availability and Claude Mythos 5.1 in restricted access on Monday, with a pricing change that immediately became the most-discussed AI pricing development since OpenAI's GPT-4 launch. Cache-read pricing for Fable 5.1 drops 75 per cent, from one dollar to twenty-five cents per million tokens. Base rates — ten dollars input, fifty dollars output per million tokens — are unchanged. Anthropic estimates the effective cost reduction at approximately 25 per cent for typical workloads and 45 per cent for highly agentic workloads that depend heavily on cached context. At the benchmark level, the improvements over Fable 5 are significant: Terminal-Bench-Science improves from 24.7 per cent to 52.6 per cent; AutomationBench from 17.1 per cent to 31.4 per cent; CursorBench from 60-something to 73.4 per cent.
The pricing structure matters more than the benchmarks for enterprise adoption. The economics of agentic workloads are fundamentally different from single-prompt use cases: an autonomous agent completing a multi-hour investigation reads and re-reads large context windows repeatedly, accumulating cache costs that can dwarf the base inference cost. At one dollar per million cache-read tokens, the TCO for agentic Claude deployments has been the primary objection from enterprise procurement teams for the past two quarters. At twenty-five cents, that objection has a materially different answer. Anthropic's estimate of 45 per cent cost reduction for highly agentic workloads is not marketing language — it reflects the actual cost structure of agent loops.
The Mythos 5.1 deployment strategy is the more conceptually interesting development. Mythos 5.1 runs the same underlying model as Fable 5.1 but with different safeguard configurations, available only to organisations that have passed a vetting process focused on cybersecurity and life-sciences applications. The rationale is not explained in detail by Anthropic, but the implications are legible: Mythos 5.1 is designed for use cases where the maximum capability of the model is required and the operator can be trusted to manage the risk surface. Life-sciences organisations are deploying it for protein design; one result reported in the release materials is experimentally validated protein binders produced by the model's autonomous research capability.
The dual-release structure attempts something that the frontier AI field has struggled to operationalise: deploying different capability profiles to different classes of operator based on trust level, rather than releasing a single version and relying on post-deployment content policy to constrain misuse. Whether it works as a safety mechanism depends entirely on the quality of the vetting process, which Anthropic has not described publicly. As a product decision, it is the clearest statement yet that Anthropic views responsible deployment as a distribution strategy, not a constraint on distribution.
The practical recommendation for enterprise AI teams is direct: if your Claude workloads involve agentic task completion, long-context reasoning, or repeated retrieval over large document sets, the pricing change makes a new set of use cases commercially viable that were previously uneconomical. Run the numbers against your current token logs. The 45 per cent reduction on highly agentic workloads is not uniformly distributed — it applies precisely to the highest-cost, highest-value use cases, which are exactly the ones enterprises should be prioritising.