Build1 publisher3 min readPublished
A 90% cache-read discount takes 81% off a 10,000-token prompt's input line
A dev.to post tells small businesses to compare cache read prices before per-token stickers. Run its own example prices through its own prompt and the input-side saving comes out at 81%, not the 55% it claims.
The Engineer · Build desk
What happened
- A dev.to post models a typical business prompt at 10,000 tokens, of which 9,000 are identical on every call and only 1,000 change with the question or the customer record.
- Its example of a good vendor answer is $0.003 per 1,000 regular input tokens against $0.0003 per 1,000 cached, a cache read price one tenth of the standard rate.
- It also says the AI code editor Cursor was banned from a major provider in September 2026, leaving customers who kept both their AI memory and their files there with neither.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision The comparison to run before signing is your own prompt priced at each vendor's cache read rate, because on repetitive workloads that rate, not the headline per-token price, sets the invoice.
- contradiction The post's 55% claim and its own example prices do not agree, and the gap matters: the 81% figure covers input only, so a workload that emits long outputs will save considerably less than either number suggests.
- cost The buyers exposed are the ones without a procurement function modelling total cost of ownership, who pay the full input rate on the same 9,000 tokens every call because nobody asked the vendor for a cached rate.
Caching, in the sense this post uses it, is a billing rate. The provider keeps a copy of the repeated part of the prompt so it does not reprocess the whole thing on every call, and bills that part at a lower rate [4].
Take the post's own example prices: $0.003 per 1,000 input tokens against $0.0003 per 1,000 cached [5]. A 10,000-token prompt with nothing cached costs $0.03 on the input line [2]. Cache 9,000 of it and the same call costs $0.0027 for the cached block plus $0.003 for the 1,000 tokens that changed, or $0.0057 [3]. That is 81% off input [4], 26 points better than the 55% per call the post advertises [5]. The two figures measure different things. The 81% is input only, and the example prices the post gives are input prices [5].
The post calls the per-token price "almost irrelevant for small businesses" [1]. Its own example puts the cached rate at a tenth of the standard input rate [5], so the sticker still sets both rates.
For the 81% to appear on an invoice, those 9,000 tokens have to be the same text call after call. The post's list of repeated content includes conversation history [3]. History gains a turn every call, so that part of the block changes every time.
Once the discount lands, the small part of the prompt costs more than the big one: $0.003 for the 1,000 changing tokens against $0.0027 for the 9,000 cached ones [6]. Cutting 1,000 tokens of boilerplate then saves $0.0003, while cutting 1,000 tokens from the variable payload saves $0.003, ten times as much [7].
The checklist the post recommends asks about discount depth. It says the answer should be 70 to 90% for a caller who sends the same system prompt and business context every time [8], and it treats cache pricing that exists only on an enterprise plan as a red flag [9]. A vendor who cannot answer is "optimizing for their margin," the post says [10]. The evidence offered for the headline saving is one vendor anecdote: according to the post, Anthropic kept Claude Fable 5.1 at the same sticker price as Fable 5 while the cache read price dropped, cutting real-world costs roughly 45% for heavy users who send consistent context [7]. The post does not cite Anthropic's price list.
The portability argument rests on a separate incident. The post says the AI code editor Cursor was banned from a major provider in September 2026, and that thousands of customers who had one company holding both their AI memory and their AI files were left with neither [11]. Its remedy is to keep system prompts and business rules in a document you own, keep data exportable, and keep the workflow replicable across at least two providers [12]. If the context lives in your files, you can price the same 10,000 tokens against a second vendor's cache rate before moving anything.
What to watch
- Whether Anthropic's published cache read prices for the two versions the post names support the roughly 45% cut it attributes to them.
- Whether vendors keep cache discounts on paid tiers only, the red flag the post tells buyers to test for.
- Vendor disclosure of cache write charges and how long an entry survives, both of which decide whether the 81% input saving holds across a real call pattern.