Usage and costs | Sovereign AI Docs

Organization Settings > Spend is four tabs in order: Usage meters infrastructure (agent awake time), Costs meters AI tokens by provider and user, Productivity grades what the spend produced, and Limits enforces the budget. This chapter walks them in the order the page does.

Usage: the infrastructure meter

The first tab meters infrastructure. Per-user agent awake time and sleep patterns: the org pulse chart, total awake hours with awake-now presence, wakes and sleeps in the window, and a weekly rhythm heatmap of when instances are actually on. Lenses switch between Hours, Cost (awake time priced at the org compute rate), and Value (usage attributed to swarms, flows, and chat). Auto-sleep does the saving and the page quantifies it: idle compute reclaimed, dollars saved, and how much of the work the policy did on its own. Sleeping instances are deallocated and accrue nothing.

The full Usage tab, top to bottom: org pulse, awake totals, auto-sleep savings, weekly rhythm heatmap, and the per-user table with a row expanded into tokens, sessions, and wake intervals — ↳The whole meter: pulse, rhythm, and every user row - one expanded into its token and session detail.

Costs: the AI token meter

The second tab meters AI tokens. API cost by user, project, model, and flow: pick a time range from 1 hour to 90 days or all time, filter by project or user, and review the trend, the provider breakdown, the cost matrix, the largest individual requests, a month-end forecast built from month-to-date spend, and suggested cost optimizations. Tables export to CSV, and a banner surfaces projects that are over or approaching budget.

The full Costs tab: 30-day total with month-end projection, provider breakdown, cost matrix, largest requests, and optimization suggestions — ↳The token ledger end to end: projection, providers, the cost matrix, and where to save.

How pricing works

Costs and swarm estimates price per million tokens from the live model catalog; a static table backs any model the catalog has not priced, so an unpriced row never zeroes an estimate. Self-hosted models are the deliberate exception: they price at zero because the org runs the hardware, and that zero is kept rather than falling through to an API-rate default. Swarm cost estimates combine catalog pricing with per-role token profiles (an architect, a backend dev, and a code reviewer consume differently).

Productivity: the outcomes lens

The third tab grades what the spend produced: spend quality (efficient, productive, waste, unclassified), the waste ledger itemizing where the red bucket came from, value by department, and cost per shipped outcome. It gets the full treatment in its own chapter.

The Productivity tab: spend quality bands, the waste ledger, and value by department — ↳Every dollar classified by what happened to the thing it built.

Limits: the enforcement

Organization Settings > Spending Limits sets the caps. Enable the cap, choose a daily or monthly period, and set an org-wide cap plus a default per-user cap. You can add finer caps per project, per model, and per flow, and set per-department budgets for the departments defined on the Departments page.

The Spending Limits admin page with the org cap, warning thresholds, and department budgets — ↳The budget with teeth: caps, thresholds, and per-department budgets in one place.

Alerts, requests, and enforcement

Usage bars turn to warning at 80 percent and exceeded at 100 percent of a cap. The alerts feed records every warning and block with the user, usage, and cap at the time. Members who hit a limit can file a cap increase request (org cap, per-user cap, project cap, or model cap); requests land here as pending for an admin to approve or deny.

Cap changes are recorded in the audit trail as spending cap change events, so budget history is reviewable alongside everything else.