Analytics and usage

Analytics and Usage give you two views of the same underlying activity: Analytics is about understanding how the project is used; Usage is about tracking spend against a budget.

Analytics

Settings › Analytics is available to any project member and requires the analytics_enabled admin flag. Data is scoped to the currently selected project and refreshed from the API, cached for a few minutes.

Choose a date range with the quick presets (Today/Last 7d/Last 30d/Last 90d) or custom From/To pickers, then switch between the tabs below. Each preset covers whole calendar days, from 00:00 on its first day to 23:59 today, and the To minute is included; for the last 24 hours or another exact window, set From and To by hand. From cannot be later than To. Daily charts group activity by UTC day, so outside UTC one local day can appear as two chart columns. The COST tile is billed spend per billing period (a calendar month), so it names the period it covers; under Today it is the month to date.

Overview

Team size, AI adoption, LLM calls, tool runs, and the top adopters.

Costs

Estimated LLM spend by day, by user, by model, by agent, and by tool.

Tokens

Input, output, and cache token usage by day, user, and model.

Agents & Pipelines

Run counts, latency, and errors per agent or pipeline.

Tools

Call counts, latency, and errors per tool.

Users

Per-user activity, adoption, and the tools and agents they used.

Health

Request volume, error rates, and latency by event type.

Select Export to Excel to download all tabs for the selected range as a single workbook.

What counts as an LLM call

The LLM calls, tokens, active users and cost figures count completed model calls only. A model call is a request to a model route: chat, completions, embeddings, responses, messages, images, audio or realtime. A completed call is one the gateway answered with a status below 400. A budget refusal, a rate limit, a validation error or a provider failure is not counted, because the Usage tab does not bill it. The Agents & Pipelines tab counts every model-call attempt, because it shows errors. The Health tab counts every request on every route, including model listings and requests to unknown paths, because it exists to show what failed. Analytics reads the gateway request log, which keeps 92 days of requests.

Cost by agent and by tool is an estimate, not the accounted spend

The Costs tab’s Cost by Agent and Cost by Tool tables answer which agent or tool a call is attributable to (#875), the same estimate the by-model and by-user tables already give — derived from the request log and the price catalogue, not the figure the Usage tab’s budget meter accounts. Cost by agent covers calls made from a runtime execution, which most direct /llm traffic is not; cost by tool attributes an execution’s whole LLM spend to every tool that execution called, so a tool used inside a longer turn is not “what that one tool call cost” in isolation. Both tables are absent — not a zero row — for a deployment or a window that cannot correlate either dimension yet. Setting a spend LIMIT per agent or per tool is not available; see Budgets and LLM Governance.

Usage

Settings › Usage requires the cost_budgets_enabled admin flag and shows spend against the current monthly billing period for platform-shared models only — models configured with your own provider credentials are billed by that provider and are not included.

The page shows, for the current billing period:

• Spent this period, Limit (or Unlimited) and Remaining.
• A usage meter with the percentage of the limit used. It changes colour past the project’s warning threshold and again at the limit; a scope with no enforced limit says so instead of showing a meter.
• The billing period’s start and end dates.

A member who is not a project administrator sees the percentage and the period but not the amounts. When a request is refused because your own member budget is exhausted, the error links to this page scoped to you, showing your own spend against your own limit in this project.

A warning banner appears as usage approaches its limit; once the limit is reached, requests to platform-shared models are refused until the limit increases or the period resets. Select Refresh data (the circular-arrow button in the header) to fetch the latest figures without reloading the page; if a refresh fails, the last loaded figures stay on screen with a note that they could not be updated.

Troubleshooting

Analytics or Usage section is missing

The relevant admin flag (analytics_enabled or cost_budgets_enabled) is off for this deployment — ask your administrator.

Numbers look stale

API results are cached briefly; wait a moment, or select Refresh data on Settings › Usage.

Requests are refused with a budget error

Your project or personal budget for shared models has been reached for the current period — wait for the reset or ask an administrator to raise the limit.

Related

Admin: Budgets and governance · Project settings