Governance usage & cost dashboard

Organisation admins can read a live cost and usage ledger for every AI capability call made by their organisation — broken down by team, workspace, user, feature, model or, for the AI Gateway, principal and tool — directly from Settings → Governance → Cost Insights.

---

Opening the dashboard

1. Sign in as an organisation admin.
2. Open Settings and select Cost Insights under the Governance section in the sidebar.

The page opens on the Usage tab for the current month, grouped by team. Three tabs cover the page:

Everything you select lives in the URL (?tab=…&range=…&source=…&measure=…&by=…), so a filtered view can be shared or bookmarked.

---

The budget line

Under the page title, on every tab, a strip shows the organisation's spend cap for the current cap period (the month, or the day when caps are daily): spend so far, the cap, a dashed tick where the current run rate lands at the end of the period, the share of the cap used, and the days left. It always describes the cap period — changing the range below it does not move it.

---

Range and source

Two controls in the page header apply to the Usage and Pricing tabs:

Changing either re-reads the ledger on the server before the page re-renders. Choosing AI Gateway as the source scopes every number to router traffic (see AI Gateway) and swaps the breakdown's dimensions to Principal and Model — a principal is a person or a service account.

---

Usage

The four measures

The row of headline figures is also the switch: click Spend, Calls, Tokens or CO₂ and the chart and the breakdown re-project in that unit.

• Spend — ~$1,204.18. The tilde marks an estimate; the hint opens the pricing explanation.
• Calls — exact count, with the success rate and the number of failed calls.
• Tokens — exact count, split into input and output.
• CO₂ — ~6.10 kg, with the equivalent distance driven.

Estimated figures (spend, CO₂) carry a leading ~ everywhere on the page; exact counts (calls, tokens) never do.

Spend per day

One column per day in the selected measure. Weekend days are drawn lighter; the day the range peaked on is named in the chart's header. Hover a column for the exact value.

Breakdown

A ledger table with one row per group, ranked and shared on the selected measure, with Spend · Share · Calls · Tokens · CO₂ · vs previous columns. The comparison is against the window of the same length immediately before the range. Rows show names for team, workspace and member. A service account currently shows by its client id rather than a resolved name; a row with no captured principal — written before this feature shipped — shows as Unattributed.

The dimension control on the table switches between Team · Workspace · User · Feature, or Principal · Model when the source is the AI Gateway.

Export CSV downloads the table as usage-by-<dimension>.csv with the columns group, cost_usd, share_pct, calls, tokens, co2_kg. The button is disabled while the table is empty.

By capability, model and workflow

Three compact lists under the breakdown: consumption per capability (token-based capabilities report token counts; unit-based ones — speech, OCR, image, video — are metered by units), the top models by spend, and the top workflows by runs. With the AI Gateway as source, the first list is instead the tools reached through the MCP endpoint with their call counts and share.

When the range is empty

The page says so — No usage for Aug 1 – 7 — explains that every model turn, tool call and workflow run lands here as it happens, and offers Show this month.

---

Limits

Spend caps

Caps are evaluated against the usage ledger at admission time, so they hold across every replica. The form edits three scopes:

Each scope takes a spend cap (USD) and a calls cap; blank means uncapped. Above the rows, Period (monthly, resets on the 1st; or daily, resets at UTC midnight) and When a cap is reached (refuse further calls, or record it and allow the call) apply to every scope.

Beside every scope, a meter shows this period so far against the cap — the organisation's own spend, or the highest consumer for a per-team or per-developer cap — with the same dashed run-rate tick as the budget line. The meter measures exactly what admission measures: every metered call since the period started, on the indicative spend. Amber from two thirds of the cap, red at it.

When a call is refused, the developer's tool receives a rate-limit error that names the cap, and Retry-After is the real time until the period resets. An administrator raises the cap here.

Router guardrails

The DLP posture on router traffic select applies data-loss prevention to developer traffic only: inherit the organisation policy (the default), enforce (block a turn that trips a rule), or audit (record the finding, allow the turn). Chat, workflows and every application surface keep the organisation policy. See AI Gateway for why source code needs its own posture.

---

Pricing

Where the spend was priced from

A stacked bar splits the range's spend by the source of the rate that priced each call — catalog rate, your override, or the default rate (and unpriced when a call carried no rate at all). Under it, Priced at the default rate this period lists the models that fell to the default, with the spend they carried and the rate applied; Set price adds a prefilled row to the overrides below. When every call was priced from the catalog or an override, the list says so.

Model price overrides

Pick an installed model from the dropdown (grouped by capability) and set its price. Token-priced models (LLM, embeddings) take a $ / 1M tokens input rate (LLM also takes an output rate); unit-priced models take a $ / unit rate. A dated deployment such as gpt-5.4-2026-03-05 automatically reuses the base gpt-5.4 rate. A cell that does not apply to the capability shows a dashed —.

Capability defaults

A fallback rate per capability (LLM, embeddings, speech-to-text, text-to-speech, OCR, image, video, web search, moderation), used when neither the integration catalog nor a model override prices a call. Leave a field blank to keep the built-in indicative default, shown as the placeholder. This is how capabilities without an enumerable model list — OCR, image, video — are priced.

---

Permissions

The underlying client.billing.* API requires an org-admin session. See Billing and usage metering for the full API reference, including the period-usage and pricing-sources reads this page uses.

---

Data freshness

Usage events are written by the capability runtime at call time. The dashboard queries are live — there is no scheduled aggregation delay. Events from capability calls that occurred within the selected date range are included immediately.