Anthropic's new flagship costs exactly what the old one did. The whole discount is hidden in the prompt cache

A new top-of-the-range model that does not raise its price is unusual enough; one whose maker cuts a bill by up to 45 percent while leaving the headline rate untouched is rarer still. Anthropic released Claude Fable 5.1 on 1 September 2026 at $10 per million input tokens and $50 per million output — the same figures Claude Fable 5 carried since June. The change sits one line further down the price list: reading from the prompt cache fell from a tenth of the input price to a fortieth, or $0.25 per million tokens. That line matters more than it looks. In agent work the same instructions, tools and files are re-sent on every step of a long task, and it is the cached re-reads, not the fresh input, that pile up. Anthropic puts the saving at roughly 25 percent for typical workloads and up to 45 percent for heavily agentic ones — an effective price cut delivered without touching the number buyers compare. On capability, the maker's own figures put the biggest jump in scientific agent work: Terminal-Bench-Science 0.1 rises from 24.7 percent for Fable 5 to 52.6 percent, more than doubling. The model reports 73.4 percent on CursorBench 3.2.0 and 65.0 percent on Humanity's Last Exam with tools. Its context window (1M tokens) and maximum output (128K) are unchanged, and its knowledge runs to June 2026. One design decision sets it apart from the rest of the line-up: adaptive thinking is always on and cannot be disabled. Every answer is a deliberated one, and the maker openly labels the model's latency as the slowest it sells — which is why it does not recommend it as the default. That role belongs to Claude Opus 5 at half the price; Fable 5.1 is for demanding reasoning and long-horizon agentic work, or for cases where Opus 5 at high effort still falls short. Developers moving over should note three breaking changes: forced tool use now returns an error, earlier models cannot read Fable 5.1's thinking blocks, and editing an earlier turn in a conversation invalidates them. The same day, Anthropic shipped an identical model under a second name. Claude Mythos 5.1 shares Fable 5.1's specifications and pricing to the last figure, but is offered by invitation only through Project Glasswing, aimed at verified cybersecurity and life-sciences professionals; access is arranged through an Anthropic, AWS or Google Cloud account team, and at launch covers US organisations. The split is the same one Anthropic drew in June with Fable 5 and Mythos 5: one model with safety classifiers in front of it, one without, so that the unfiltered configuration can be tracked — and withdrawn — separately. Claude Fable 5 moves to the maker's legacy list and remains available.