Analytics

Analytics aggregates trace data into usage and cost breakdowns — the "how much is this actually costing us" view.

What's on the page

• Total cost, total tokens, average latency, and requests/day over the last 14 days.
• A daily cost trend line.
• Cost broken down by model.
• Request volume broken down by route.

Where the numbers come from

Everything here is derived directly from the trace log — the same data that powers the Traces section — aggregated server-side rather than stored separately. There's no independent "analytics" data model to keep in sync.

Reading the cost-by-model chart

Cost isn't just about which model is cheapest per token — a "cheap" model that needs longer prompts or produces more verbose completions to hit the same quality bar can end up costing more in aggregate. The cost-by-model breakdown is the fastest way to notice that before it shows up as a surprise bill.

Requests by route

Route is one of the free-form metadata fields on a trace (see Traces). Grouping by it answers "which part of the product is actually driving LLM usage" — useful for deciding where an evaluation dataset or a tighter prompt would pay off most.

What's not here

Analytics only looks at traces — production-style calls. Evaluation run cost isn't included, since runs triggered through the real scoring engine don't call a priced model at all (see Evaluations).