A file ID is a pointer, and the pointer gets resolved before the model sees anything. The bytes stop crossing the wire. The tokens still land in the context window. That is why the upload-once arm billed roughly 3,050 input tokens per question, the same as pasting, plus about 25 tokens for the reference itself [6][7].
Run the five-question workload out. Pasting bills about 5 x 3,050, or 15,250 input tokens [1]. Upload-once bills 15,375 [2]. Caching is denominated differently, so convert to normal-input equivalents: the 2,990-token cache write at a 25 percent premium is 3,737.5, the four reads at about a tenth of input rate are 1,196, and the five questions add roughly 275 at 48 to 63 tokens each [8][9][10][11]. Total near 5,200 [3]. That is about a third of the pasted arm [4].
The break-even is early. The write premium costs 747.5 equivalents once; each cached read saves 2,691 against pasting the document again [5]. One repeat request pays for the write. That is the whole argument for the cache marker, and it does not depend on volume.
The five-minute idle expiry is what decides whether any of this transfers [12]. A support bot answering five questions in a burst holds the cache. A bot answering one question every ten minutes writes a fresh cache entry every time, which bills 1.25x what pasting would have [6]. For the 66 percent figure to hold on your workload, the traffic has to be steady inside the window, the cached prefix has to stay byte-identical, the questions have to stay short relative to the document, and you have to be on claude-sonnet-5, as the test was [3]. The New Stack's numbers are input tokens only; output was not part of the comparison.
Adoption cost splits the same way as the mechanism. The Files API upload was one call for one file ID, and the ID worked in every following request [16]. Caching required restructuring the request, moving the document into the system prompt behind a cache marker [13]. Two unrelated things broke first: temperature is deprecated on claude-sonnet-5 and the API rejected temperature zero, and one response led with a thinking block that crashed the print path. Both fixes were one-liners [14][15].
The accuracy result deserves a note, because boring is the correct outcome here. All fifteen answers matched the key across all three arms, including the details planted to be missed, such as token scopes that cannot be edited after creation and the separate rate limit on the bulk endpoint [5]. Delivery mechanism did not change what the model knew.
So pick by the problem you have. The Files API is document lifecycle. Prompt caching is the token bill [17].
