Acuvo · measured, not marketed
Which provider route is honest about its cache
Everyone publishes which model is smartest. Nobody publishes which provider route actually caches your prompt, because that means paying several of them for the same real work at the same time and metering every call. We run a multi-provider chain and meter every call, so we can. A cached prompt token bills at about a tenth of a fresh one, which makes this the number that decides what a token costs — and it is on no pricing page.
Read this as a small sample, because it is one. Acuvo has no paying customers yet, so every request below is our own real work — building software through our own product — not a benchmark and not a customer’s traffic. It is honest and it is thin. Rows under 25 requests are shown and labelled rather than hidden, so this table cannot quietly become a list of our best results.
OpenInference cached 83.4% of the prompt tokens we sent it. Makora cached 74.7% — a 1.1x spread. Same gateway, same weeks, same code. Neither provider publishes this number, and it is the one that decides what a token actually costs.
- metered requests
- 1,000 of 10,066
- prompt tokens
- 20.3M
- output tokens
- 0.78M
- window
- 2026-09-18 → 2026-09-20
This is the most recent 1,000 of 10,066 metered requests, not all of them — the database caps a single read and reports no error when it does. The dates above are this window’s, not the whole record’s.
By upstream — the leg that actually served
| route | requests | prompt tokens | cached | cold starts | standing |
|---|---|---|---|---|---|
| Relace | 662 | 13.45M | 80.4% | 25.8% | measured |
| Wafer | 106 | 2.19M | 83.3% | 16.0% | measured |
| Makora | 59 | 1.64M | 74.7% | 23.7% | measured |
| (not an OpenRouter call) | 67 | 1.35M | 63.8% | 34.3% | measured |
| OpenInference | 83 | 1.20M | 83.4% | 0.0% | measured |
| DeepInfra | 17 | 0.30M | 60.5% | 47.1% | too thin to publish (under 25 requests) |
| CoreWeave | 2 | 0.07M | 0.0% | 100.0% | too thin to publish (under 25 requests) |
| StreamLake | 3 | 0.06M | 36.3% | 66.7% | too thin to publish (under 25 requests) |
| Sail Research | 1 | 0.01M | 0.0% | 0.0% | too thin to publish (under 25 requests) |
By model
| route | requests | prompt tokens | cached | cold starts | standing |
|---|---|---|---|---|---|
| deepseek/deepseek-v4-flash-0731 | 987 | 20.18M | 78.6% | 23.9% | measured |
| deepseek/deepseek-v4-flash-0731:free | 12 | 0.10M | 77.6% | 0.0% | too thin to publish (under 25 requests) |
| deepseek/deepseek-chat | 1 | 0.01M | 0.0% | 100.0% | too thin to publish (under 25 requests) |
What we could not measure
- 1 of 1,000 requests carry no cost at all, because we could not price the model against a rate we trusted. They are recorded as unknown rather than as free. An earlier version of this meter guessed instead — it fell back to the dearest rate in the table and booked a free-tier call at $0.56 — and a guessed number is worse than a missing one, because a missing one is visibly missing.
- 20 requests came back with no cache figure. Silence is counted as unknown, never as a miss: an upstream that stops sending the field would otherwise look exactly like a cache collapse. Across the whole sample, 98.0% of requests answered the cache question at all.
- No latency here. We do not yet record how long a provider took to answer, so we publish nothing about speed. The call already happens; the clock is simply not read. Saying that is cheaper than implying we measured it.
Privacy
We publish no prompt, no completion and no customer identity on this page — only counts and rates about models and providers. We cannot make the wider promise: some providers we route to, DeepSeek among them, retain user inputs under their own terms, and that is their server and their policy, not something we control or can undo by writing it here.
Measured through acuvo-code, our CLI — npm i -g acuvo-code. Every figure on this page is a count or a ratio over our own metered calls, recomputed on load.