Your AI spike happened three days ago. Your dashboard just told you.

01 — YOU SEND THE COUNTS

One call, after your request has already returned

Your provider hands you a usage object with every response. You pass those counts to track() with a feature name. That's the integration.

It is fire-and-forget and sits outside your request path, so it cannot slow down or break what your users are waiting for. Prompts and completions never leave your infrastructure — we take counts, not content.

02 — WE PRICE EACH COMPONENT

Seven components, seven rates, never blended

A modern request is not two numbers. It is text in, cached reads, cache writes, audio in, reasoning, text out and audio out, and each one bills at its own published rate.

billableInput  = prompt_tokens − cached − audio_in
billableOutput = completion_tokens − reasoning − audio_out

cost = billableInput  × input_rate
     + billableOutput × output_rate
     + cached         × cache_read_rate
     + cache_writes   × cache_write_rate
     + audio_in       × audio_in_rate
     + audio_out      × audio_out_rate
     + reasoning      × reasoning_rate

03 — CACHING COUNTS PROPERLY

Cached tokens bill at the cache rate, not the input rate

If you cache a prompt prefix, those tokens cost a fraction of full price — 90% less on one flagship model, 50% less on another. The discount is never a constant, so we read each model's own cache rate instead of assuming one.

Providers also disagree about whether cached tokens sit inside the prompt count or beside it. Get that backwards and you charge them twice. We resolve it per request, checking the model before the provider, so a model served through a reseller still resolves correctly.

Cache writes cost more than ordinary input. We show that too, because caching can make a bill worse and nobody warns you about that.

04 — YOU SEE IT IN UNDER A SECOND

And we tell you when we had to approximate

Rates come from each provider's published schedule and are re-synced every day, so a price cut appears without anyone hand-editing anything. Cost is computed on our side, at ingest — we never accept a cost figure from your client.

When a provider publishes no rate for a component, we fall back to the standard multiplier and mark the row as derived. When we can't tell which convention a response follows, we price what we're sure of and flag the rest. A number you can see is uncertain beats a number that's quietly wrong.