A report is always scoped to one provider. OpenAI and Anthropic report cost differently, so their reports make different claims — see Per-provider differences below.
Opening the report
From the agent’s Overview section, use the Audit report buttons next to the provider tabs. On a provider tab, the button opens that provider’s report. On All providers, each provider with usage gets its own labeled button, so you always know exactly which provider’s report you’re opening. The report opens in a new browser tab as a standalone page. An optional Report window start/end date pair next to the buttons scopes the report to exact dates. Leave both blank and the report follows the date range you’re viewing on Overview. The buttons appear once at least one connected provider has produced usage. To keep a copy, save the report as a PDF from your browser’s print dialog. The report has a print-friendly layout built in, and its print styles keep full colors, so the saved PDF matches the report on screen and can be forwarded as-is. If the agent hasn’t reconciled any usage for that provider yet, the report renders a “No audited spend yet” state instead of fabricating numbers. Run a sync and audit first — see Sync now and Refresh insights.What the report contains
The report opens with a masthead (customer, provider, and the audited window) and a KPI band, then walks through seven numbered sections.KPI band
Up to four headline numbers across the top:- Audited spend — Total billed cost across the reconciled window shown in the report header.
- Growth — Spend growth from the first month to the last complete month, compared at per-day rates so a partial first or current month can’t distort the figure.
- Biggest month — The highest-spend complete month. A partial in-progress month is never counted as a record.
- Recent pace — Average spend per day over the trailing window (up to 14 days). This same trailing window is the basis for the annualized savings estimates, so the two never disagree.
1 · Executive summary
The paragraph a CFO reads when they read nothing else: what was spent, where it’s heading, whether the bill checks out, what caching already saved, and the single largest action on the table. Everything it claims is expanded in a numbered section below.2 · Monthly spend — billed vs expected at list
A month-by-month table of billed cost against the same usage re-priced at the provider’s published list rates, with a difference column and a totals row. A positive difference means the provider billed above its published list price for that usage; negatives are effective discounts. A month clipped by the window edge is marked as in progress. Differences large enough to matter become the itemized findings in section 4. This table is what makes the billing-integrity statement checkable line by line.3 · Spend composition
A plain-English narrative that walks through monthly spend, the current month’s daily pace, and weekend-vs-weekday mix (steady weekend spend suggests production workloads; a weekend drop suggests interactive use), followed by a model-mix table showing each top model’s spend and share of the bill.4 · Findings that need an owner
The highest-impact open findings as numbered rows, each with its full explanation and one honest per-period dollar figure. Multi-day figures are normalized to a 30-day rate so items are comparable; single-day events state the day; each figure covers only the finding’s own measured period. Duplicate instances of the same rule are collapsed to the highest-impact one. The report also calls out new model adoption: models that first appeared partway through the window and are already running at a material annual pace, so you can confirm the rollout was deliberate.Finding figures run on different time bases, so they are deliberately not summed into the savings headline. Each carries its own period label.
5 · Savings
Realized savings come first: when prompt caching saved money in the period, the report states the amount (cached reads priced against full-rate input) and labels it as money not spent, not a projection. Modeled opportunities follow, with a combined headline range at your current usage:- Batch pricing — Provider batch APIs are a flat 50% off for async-tolerable work. The report estimates this from your annualized embeddings line and only shows the item when the saving is material.
- Negotiated rates — At list price with a sufficient annual run rate (above $100,000 a year), a 10–20% discount against a 12-month commitment is commonly reported market practice. The report states this as a market benchmark, not a computed entitlement.
6 · Billing integrity
A statement of what the daily reconciliation confirmed for this window: whether billed totals reconcile against provider-reported usage at the account level, and whether any billing discrepancy is confirmed. Anything confirmed links into the Recovery flow.7 · Methodology & definitions
The definitions behind every figure above, so a reviewer can audit the report itself: how audited spend and expected-at-list are computed, why growth uses complete months at per-day rates, how finding figures normalize to 30-day and 365-day rates without ever summing across incompatible periods, when per-model figures are apportioned estimates, and the data-access statement (read-only admin key against the provider’s reporting APIs only; prompts and outputs are never read; audits run daily).Per-provider differences
The report says different things per provider because the providers report cost differently, and the report only makes claims its data can back:- OpenAI — OpenAI’s cost API is model-blind: it reports cost per project, so per-model billed cost is an apportioned estimate. An OpenAI report therefore never claims a confirmed overcharge. Its integrity statement stays at “no material provider overcharge confirmed,” and any recovery is invoice-side — discrepancies must be validated against your provider invoice first. This matches the Confirmed vs. Potential split in Recovery.
- Anthropic — Anthropic reports real per-model cost, so its report can state per-model reconciliation and can surface a confirmed billing discrepancy when one exists.