Anthropic released Claude Fable 5.1 on September 1, 2026. The model keeps the same base price as Fable 5: ten dollars per million input tokens, fifty dollars per million output. What changed is cache reads.
Cache reads happen when the model rereads context it already processed. Code, documents, chat history. Cache reads now cost 75% less: down to twenty-five cents per million tokens.
For typical work, Anthropic says that cuts your total bill about 25%. For coding agents that run for hours and reread the same context hundreds of times, the cut reaches 45%.
The catch lives in the architecture
That 45% only works if you design your prompts to maximize cached context. Append-only conversation history. Static system instructions. Tool definitions that never change. All of it structured to let Claude reuse what it already read.
If you switch models, you rebuild the whole setup. Every competitor structures caching differently. The savings lock you tighter to Anthropic's stack than the headline price ever did.