The cache-read column in the last four weeks of Fable usage, not the figure in the announcement, sets the saving a team actually books.
The rate on that line went from $1.00 per million tokens to $0.25, a cut of 75% [4][13]. Every other cell held [5]. So the fall in your total bill is 0.75 times whatever share of the bill cache reads were. To land Anthropic's typical figure of about 25%, cache reads had to be roughly a third of your Fable spend, since 0.25 divided by 0.75 is 0.33 [14]. To land the up-to-45% figure, roughly 60% [15]. Anthropic describes the agentic case it measured as context-heavy, tool-heavy work in which cache reads make up most of the cost, which is the same statement reached from the other end [12].
Teams often assume that using prompt caching means getting the discount, but the bill works differently. In the shape Anthropic labels long agentic sessions, a large stable prefix is re-sent every turn and grows, so cache reads dominate by construction and the discount lands nearly in full [18]. In mixed production traffic, a cached system prompt plus retrieved context sits alongside a lot of generated output, cache reads are a real but minority share, and the output rate is untouched [20]. For short one-shot calls with no reusable prefix, nothing on the sheet changed, so the saving is zero [16].
Inside Anthropic's own lineup the ordering inverts. Opus 5 charges $0.50 per million for a cache read, twice what Fable 5.1 charges, while costing twice as much on input and output [9][17]. Anthropic's migration guide puts that in front of teams moving up from Opus 5 [10]. The 2x-the-price shorthand stops being a single number and becomes an estimate about session shape, because the longer the session and the larger the prefix, the more of the bill sits on the one line where the flagship is now the cheaper model. Finding the crossover takes your own turn counts and prefix sizes.
There is a second, less visible cost. Anthropic prices cache reads at 10% of base input on every other current Claude model, and OpenAI and Google publish the same multiplier [4][6]. Meta's Contributor tier goes lower at 2.0% in exchange for training permission, and DeepSeek works out near 3.2%, but 2.5% is the lowest ratio a US frontier lab has published [7]. Any forecasting spreadsheet with input divided by ten hard-coded now returns a cached rate four times too high for this one model [8].
The forcing function is one division. Split four weeks of Fable spend into cache reads, fresh input and output, multiply the cache-read share by 0.75, and read the answer as the discount you will book. Above a third, the release is worth the migration on cost grounds alone [14]. At a tenth, the saving is 7.5% and you should be moving for some other reason [19].
