Alright, let's cut through the marketing. Every LLM cost-tracking tool promises granular cost-per-customer metrics. Helicone's dashboard shows me a nice, big, useless aggregate bill from OpenAI. I need to see which of my tenancy's customers are actually burning through $50 in GPT-4 context each day.
The docs mention tags and custom properties. I've slapped `customer_id` into the headers for our proxy calls. Now I'm staring at the "Analytics" page and it's giving me total requests, average latency, and total cost. Great. Not.
So, for those who've actually done this and aren't just repeating the sales site copy:
* Is the path to actual, actionable cost-per-customer reporting buried in some SQL query in the "GraphQL" playground? Because the pre-built dashboards seem designed to avoid showing that directly.
* If I *have* to use the GraphQL interface, what's the actual query to slice cost by a custom property over a time range? And does it correctly factor in different model costs (GPT-4 vs. 3.5-turbo) per request, or is it just averaging?
* How are you handling the attribution when a single user session might fire off multiple different model calls? Am I going to be 5% off due to rounding errors on every request, making the final sum meaningless?
I'm skeptical that this works out of the box without a PhD in their schema. Prove me wrong. Show me the concrete steps, not the conceptual flowchart.
Data skeptic, not a data cynic.
Right. We hit this wall last month. The analytics page is useless for this. You do have to use the GraphQL playground.
The query you want is under the "Cost" metrics in the schema explorer. Use `groupBy: customProperty` with your property key and a `dateRange`. It does factor in the actual model cost per request, it's not an average.
Our numbers were still off by about 8% until we realized the headers weren't propagating to our async batch jobs. Check your background processing.