Is the backend rounding up to $0.01? #48
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: co-op/support#48
Reference in new issue
Block a user
What do you need help with?
I was testing the audio STT API. My tests were very short, like "hello hello hello". I noticed that the first request took up $0.01, which seemed like a lot (it's $0.05 / 1M input tokens, and I definitely didn't go near 20% of that). I thought it could just be the UI rounding it up, not the backend, but I tried it with a second request, and now $0.02 has been used. I am used to other providers where they charge in many decimal places. (See attached files for both.)
I know, this looks trivial, but over the course of time, it adds up. If you use many small requests (each much cheaper than $0.01), you end up paying much more due to rounding errors.
What were you trying to do?
I was trying to use the audio STT API, hoping that the request would be charged with many decimal figures to maintain precision.
What happened instead?
It seems that it rounds up per request.
What have you already tried?
Tried a second time to make sure it's not just a UI rounding.
Environment
Hi,
Thanks for the careful testing — you did everything right, and the answer is stranger than rounding.
Nothing is being rounded. For
tinfoil/whisper-large-v3-turbo, Tinfoil's actual price structure is $0.01 per request (flat) — not per token. So a two-second "hello hello hello" genuinely costs a cent, passed through at cost, and your two requests totalling $0.02 is the correct sum.The wrong part is our docs. The $0.05/1M input-token figure you reasonably used comes from a stale price row in our published model list, which described a different pricing structure than the one Tinfoil actually charges for this model. We're correcting it — thank you for surfacing it.
What this means for heavy use of small requests: your arithmetic is right — a per-request floor adds up, and many small calls cost more than the equivalent audio minutes. If that matters for your use case, we're looking at adding Tinfoil's token-priced audio model as an alternative, where tiny requests cost fractions of a cent. Also a heads-up: the dashboard rows for these show "$0.01 · 0 tokens" — odd-looking but correct, since Tinfoil returns no token usage for audio.
Appreciate the report — this is exactly the kind of billing detail worth getting right.
Both changes are live:
The docs were wrong, as you worked out.
tinfoil/whisper-large-v3-turbois $0.01 per request, flat — zero per-token — passed through at cost. The models page now says so instead of the stale $0.05/1M row you priced from, and the odd-looking dashboard rows ($0.01 · 0 tokens) are explained there too.The token-priced alternative is served:
tinfoil/voxtral-small-24b— $0.20/1M input + $0.60/1M output tokens, no per-request charge. On/v1/audio/transcriptionsa short request bills ≈$0.0001: Tinfoil tokenizes audio in ~30-second chunks, so the smallest requests cost a fraction of a cent (~0.008¢) rather than a linear slice of one — about 1/125th of the whisper floor. Through the chat API withinput_audioit meters exactly per token, which is also the better route for audio longer than ~30 s. Both are on the models page with examples.Thank you — your report is what drove both changes.
Update and closing: both follow-ups are live.
tinfoil/whisper-large-v3-turbois documented as $0.01 per request (flat), and the model list now notes that Tinfoil audio rows show "$0.01 · 0 tokens" because Tinfoil returns no token usage for audio.tinfoil/voxtral-small-24b($0.20/1M in, $0.60/1M out, no per-request fee) — so heavy use of tiny STT requests no longer carries the per-request floor.greenpt/qwen3-embedding-8bandtinfoil/nomic-embed-textare callable on member keys.Your report drove all three. Closing here — if the billing still looks off on your account, reopen and we'll dig in.