Guide
Every AI bill in one place: how to track AI spend across providers
Published
If your company pays OpenAI, Anthropic, a cloud provider, a gateway and a few coding tools, no single console shows what AI costs you. This guide shows where each bill comes from, how to add them up without counting anything twice, and how to check metered usage against what you were billed.
Why is AI spend hard to see?
Because it arrives in several shapes, each in a different console: API invoices from model providers, spend reports from gateways like OpenRouter, usage from coding tools like Claude Code and Codex, and seat licences. No one bill has all of it, and some of it overlaps: a coding tool on an API key is already part of that provider's invoice.
Token Meter's AI Bill (Analytics > Bill, in team workspaces, for owners and admins) brings these into one total per date range, and keeps each source visible beside it.
Where does each provider's billing data come from?
From four kinds of source: the provider's settled invoice, an account report, a CSV you upload, or Token Meter's own metered estimate. Provider-reported billing is stored separately from metered usage and never merged with it.
| Source | Providers | How it arrives |
|---|---|---|
| Provider invoice (settled billing) | OpenAI, Anthropic, Google, Bedrock, Azure OpenAI | OpenAI's Costs and Usage API and Anthropic's cost report (Admin API keys), Google's Cloud Billing BigQuery export, AWS Cost Explorer through a cross-account role, Azure Cost Management through a service principal. |
| Account report | OpenRouter, xAI, Vercel AI Gateway, Fireworks; also Mistral and DeepInfra | The first four report spend in US dollars, which counts toward the total. Mistral and DeepInfra send reports with no dollar amount; they are kept but never added. |
| Manual bill import (CSV) | Any provider, including Together AI and a LiteLLM gateway | A file you upload. Its rows count as that provider's invoice. |
| Metered estimate | Every provider you send usage for | Token usage seen by the SDK, the router or OpenTelemetry, priced at your configured model rates, including private rates. |
You connect an account under Settings > Providers. Token Meter checks the credential with the provider before saving it, stores it encrypted, and queues the first import. For a provider with no connection, upload a CSV with at least billing_day, sku and amount_usd:
billing_day,sku,amount_usd,model_id,token_class,quantity
2026-10-01,Input tokens,12.40,llama-3.3-70b,input,15500000
2026-10-01,Output tokens,8.10,llama-3.3-70b,output,2700000The upload is previewed first, with invalid rows and their reasons, and nothing reaches the bill until you confirm. Re-uploading a row that was already imported updates it rather than adding it twice.
How do you avoid counting the same spend twice?
Count each provider from one source only: its invoice, else its account report, else the metered estimate. Token Meter never counts a provider from two sources for the same day, and it never adds metered spend on top of an invoice for the same provider.
- An invoice covers the whole range. If a provider has any invoice line in the range, its invoice total is used, even for recent days the invoice hasn't reached yet.
- Account reports count whole periods. Only report periods that lie entirely inside the range are added; a period is never prorated. Days the report doesn't reach yet are filled from metered spend, and that part is shown separately.
- Coding tools inside an invoice are disclosed, not added. Claude Code on API keys is billed by Anthropic, Gemini CLI by Google, and Codex on an OpenAI API key by OpenAI. When that provider's invoice is imported, the tool's spend is already in it, so it is shown as a disclosure line instead of being added again.
One overlap remains: Claude Code and Codex calls sent through Token Meter's router are metered too, so they can appear in both metered and coding-tool spend.
What goes into the total?
The total for a range (this month, previous month, 30 or 90 days, in UTC days) is the sum of seven lines, each shown separately so you can see where every dollar came from:
| Line | What it adds |
|---|---|
| Provider invoices | The invoice lines of each provider that has any in the range. |
| Provider account reports | For providers with no invoice but a US dollar report: the report, plus metered spend on days it doesn't cover yet. |
| Metered, no invoice or report | The metered estimate for every other provider. |
| Coding tools, pay-as-you-go | Pay-as-you-go coding-tool dollars, unless the provider that bills them has its invoice imported. |
| SaaS seats, run-rate | The current monthly seat run-rate from Licenses × days in the range × 12 ÷ 365. A run-rate, not billed seats. |
| Commitment overage | Overage recorded on a contract that is billed separately from the provider invoice. |
| Credits and adjustments | Net credits and adjustments dated in the range, subtracted. |
Illustrative numbers
A made-up 30-day range, to show how the lines combine.
- OpenAI's invoice lines total $6,400. That is used; OpenAI's metered estimate is not added.
- Anthropic's invoice is $3,100. Claude Code on API keys cost $800 of that, so it is disclosed, not added.
- OpenRouter's report covers $900, plus $40 metered on the days it hasn't reached yet: $940.
- Together AI has no connection or CSV, so its metered estimate counts: $210.
- Seats run at $1,200 a month: $1,200 × 30 × 12 ÷ 365 = $1,183.56.
- A $100 credit is subtracted. Total: $11,733.56.
Below the total, a table shows each provider's metered spend, invoice and account report side by side, and never adds them up.
How does metered usage compare with the invoice?
Every day at 07:00 UTC, Token Meter compares metered spend with invoice lines per provider and day. Drift is the invoice minus metered spend, and a day is within tolerance when drift is at most 2% of the invoice. Each provider then gets one status for the range:
| Status | Meaning |
|---|---|
| Matched | Every reconciled day in the range is within 2%. |
| Drift detected | Invoice evidence exists, but at least one day differs by more than 2%. |
| Awaiting invoice | No settled invoice has been reconciled for this provider in the range. |
| Not comparable | There is only a US dollar account report, which isn't suitable for reconciliation. |
Common reasons the two differ:
- Settlement delays. The current day is never requested, and providers can hold back a few more recent days. An invoice that looks low this month may just not be settled; check the Invoice through column.
- UTC days. Billing days are UTC, so compare against UTC days, not your local calendar.
- Negotiated prices. Metered spend uses list prices unless you add a private rate under Settings > Providers. A start date up to 365 days back reprices usage from that day.
- Unmetered calls. Metered spend covers only the usage you send through the SDK, the router or OpenTelemetry. The SDK wraps OpenAI, Anthropic, Gemini, Bedrock and Azure OpenAI clients, plus OpenAI-compatible clients for providers such as OpenRouter and xAI; see Providers.
CSV uploads are never reconciled, and neither are Azure OpenAI subscriptions billed in a currency other than US dollars.
How often does it update?
Most connections sync about every hour. Google, Bedrock, Azure OpenAI, Mistral and DeepInfra sync once a day. Reconciliation runs once a day. None of this is real time: invoices arrive as providers settle them.
- The first import brings in up to 90 days of history. OpenRouter serves 30 completed days.
- Monthly-only reports (Mistral, DeepInfra) include only complete billing periods.
- Account reports re-read the last few days on each sync, so provider corrections replace earlier figures.
What isn't included in the total?
Four things are left out, and the AI Bill lists them under the total so nobody reads it as complete when it isn't:
- Cursor usage-based spend, which isn't collected yet.
- Expense imports. They are used to find unmanaged AI tools and would repeat bills already counted.
- API-equivalent value of subscription use: what flat-rate subscription usage would have cost at API prices. It isn't money spent.
- Reports with no US dollar amount: Mistral, DeepInfra, or a report in another currency. The provider is named in the note.
Google coverage is partial. Per-model usage comes from Vertex AI only, so Gemini Developer API (AI Studio) traffic isn't in metered estimates or reconciliation unless you send it through the SDK or an observed gateway. The Cloud Billing export may still include its spend.
Full details are in AI Bill and billing imports and Coding agents. For one provider in depth, see OpenAI cost tracking; to split the total by the customers who caused it, see LLM cost per customer.
Put every AI bill in one place
Create a workspace and connect one provider account under Settings > Providers. The first import brings in up to 90 days of history, and the AI Bill shows it next to your metered usage.