See where the money goes

The first page you see. It shows where the money went, by provider and by agent, how spending moved day by day, and which keys spent the most.

Once a provider's bill is added, it also shows the bill next to what was recorded. Money on a bill that never went through your gateway gets its own line, Only on the bill, so it can't hide.

Anything worth your attention gathers on one page: a stopped run, a key with refused calls, a model with no price yet, a bill that doesn't match what was recorded. Each item links to where you deal with it.

The bell and the Dashboard point you there, and stay quiet when there is nothing to see.

Runs lists up to twelve runs in the chosen range, each with its calls, refusals and what it cost: stopped and failed ones are kept first, then the most expensive. Filter to the stopped, failed or high-cost ones, or search them by run, agent or model.

Open a run to see each call it made, in order, with what it cost. If the run was stopped, the top of its page names the limit it hit and the call where it happened, and estimates what the refused calls would have cost.

Open a call to see what its answer changed about its price: fast mode or another tier, web searches, the long-context rate. For Anthropic, a run whose calls wrote a large conversation history to the cache again, because the gap before them was longer than the cache life, says how many calls did and what those writes cost, as an estimate.

The twelve keys that spent the most, each shown by a fingerprint, never the key itself, with how many of its calls were refused. A key's limits are held by the gateway that uses it, so a paused key is resumed there.

Put a daily limit on every key shows how to set one.

Tags are how spend lands on the right run, agent, team, project or customer. Some come free: an agent started with run is one run already, and so is each Claude Code session.

For your own code, add a header or two to each call. Team, project and customer are saved with every call and come out in the exports, so finance can split the bill by them.

Every figure carries a label saying how far to trust it.

• estimated: priced from the provider's own token counts with AxiGate's price list. Every run and call starts here.
• provider-reported: from the provider's own cost data, once its bill is added.
• invoice-reconciled: someone confirmed the month against the invoice (Close the month against the invoice).
• closed on estimate: someone closed the month with no bill imported, on AxiGate's own estimate. It is never invoice-reconciled.
• unknown: the call came back without usage numbers. Its cost is not zero; the provider's bill fills it in.

A run's own figures stay estimates even after the bill is in: the bill sets the month's total, not each call's.

A run whose calls went to the provider's own API also says who pays for them, read from the credential they went on: an API key, billed per token (a Gemini key only if its project has billing on), or a signed-in account or service account, whose dollars are what the calls would cost at list price; whether that account is billed per token depends on it (a Gemini service account's project is). It is a label: it changes no limit and no figure.