LLM token usage in Elasticsearch Service Billing integration

Elastic Stack Elasticsearch
1

Hey guys!

I installed Elasticsearch Service Billing integration but cannot find my LLM token usage. I may see it in Elastic Cloud:

image
image1038×217 6.3 KB

However, I want to see this info in Elastic integration

If it is not possible, how can I track how many tokens the specific account consumed?

Thanks!

2

It may take some time to be collected by the integration, but the data is there.

it is the type inference.

3

Hey @leandrojmp , thank you for response! Indeed, inference appeared after some time

Though, I still cannot find the wait to drilldown this usage per account

So, I want to be able address the question "which account does consume the most of the tokens?" on some graph

I found Kibana API:

POST kbn://api/agent_builder/agents/elastic-ai-agent/consumption
{"size": 25, "sort_field": "updated_at", "sort_order": "desc"}

which returns an account and token usage

However, to make it useful I need to write a little script which will execute this request every N minutes, send it to Elastic and then you this dataset to build visualizations

But maybe I missed something and such data is available somewhere else in Elastic already?

Thanks!

4

For an account-level view, I would separate the problem into collection and visualization.

First, make sure each inference request has a stable account or tenant identifier, together with a request ID, model name, input/output token counts, timestamp, latency, status, and retry information. I would avoid storing the full prompt or response by default.

Then index those usage events into a dedicated data stream. This makes it easier to build aggregations such as tokens by account, model, day, and error type. It also avoids treating a billing or inference summary endpoint as the long-term source of truth.

Before writing a scheduled collector, I would verify whether the consumption API is documented and supported for the exact Elastic version and deployment type. If it is not a stable public API, exporting the usage events at the gateway or application boundary may be more reliable.

For the dashboard, useful views would include:

  • input and output tokens by account
  • estimated cost by model
  • request count and error rate
  • retries and fallback usage
  • daily and monthly trends

This also makes it possible to enforce account-level limits before a request is sent, rather than detecting overuse after the billing data arrives.