Reading a million cached tokens back into Claude Fable 5.1 now costs 75% less, or $0.25. Anthropic left the rest of the card alone, at $10 per million input tokens and $50 per million output, and puts the saving at around 25% for typical workloads and up to around 45% for complex coding and highly agentic tasks, measured over four weeks of its own August usage.
Two days later OpenAI listed GPT-6 Astra at $10 for a million input tokens, $50 for a million output and $1.00 for cached input, the same sticker on the first two numbers and four times Anthropic's price on the third. On long-context requests the same model costs $20, $75 and $2.00. Cached reading now costs a tenth of the input price at OpenAI and a fortieth at Anthropic.
On context-heavy, tool-heavy work, Anthropic says cache reads make up most of the cost. An agent re-reads the system prompt, the repository and the transcript on every turn. The discount is aimed at that workload, and it is invisible on a comparison sheet that carries two numbers per model. For agentic work the number that decides the bill is 25 cents against a dollar.
Our own basket records list input and output only, and every Anthropic and OpenAI model in it came back unchanged. What moved was DeepSeek V4 Pro, whose cheapest listed route fell to $1.04 and $2.08 a million tokens from $1.60 and $3.20, taking the support-assistant floor to $271 a month from $416. All four neocloud GPU ranges held for a fourth week.
