If none of these is your problem, we are probably not what you need — and we would rather you found that out here.
The model provider bills you per thousand input tokens and per thousand output tokens, at different rates per model. If your own pricing is a flat monthly fee, your margin moves every time a customer changes how they prompt — and you find out at the end of the month.
At £0.000012 per token, naive float arithmetic drifts. Multiply that across tens of millions of tokens and the total on your invoice stops matching the total in your database, which is the kind of discrepancy that costs you a customer's trust rather than a rounding error.
"Why is this month £400 more?" is a support ticket you cannot answer from an aggregate. You need the events, per model, per direction, and the rate that applied at the time.
It starts as a counter in Redis. Then it needs deduplication, late events, retries, a plan that changes mid-month, and an audit trail. It becomes a permanent claim on an engineer.
Recommended: Python or Node.js SDK
The Python and Node SDKs handle auth, batching and retries, and survive a restart without losing usage. Send a token count with the model and direction as properties, and the rating engine matches on the model name — the same engine that does telecom prefixes, using a different property.
Read the setup docs| Event type | Unit | Matched on | Example |
|---|---|---|---|
| tokens | tokens | properties.model | atlas-large |
| api_call | requests | properties.endpoint | /v1/verify |
| custom | seats | properties.plan | team |
Sample AI workload for a fictional customer, with input and output tokens rated per model.
Nebula Labs
LiveMonth to date
£38,863.54
Events
7,345,209
Busiest day
£2,264.82
Usage this week
Spend by type
Sample data for a fictional ai & saas business · 75M units metered this month
Click around the full version in the sandbox — no signup, no card.
One command, or one call from the system you already run. The sandbox is open if you would rather look first.