An AI provider's usage export is its machine-readable record of consumption: an API that lists the units used, such as tokens, requests or images, per time bucket and per project, workspace, API key or model; the cost follows from it. It matters for the bill because models bought directly from OpenAI or Anthropic appear in no cloud export, while the same models bought through a cloud land on that cloud's invoice and in its export, with other units and other keys.
How it works
OpenAI
An organisation's Admin API, called with an admin key, has a usage endpoint per kind of use, among them completions, embeddings, images, audio, moderations, vector stores, file search and web search. The completions endpoint returns buckets of 1m, 1h or 1d with input_tokens, output_tokens, input_cached_tokens and num_model_requests, groupable by project_id, user_id, api_key_id, model, batch, service_tier and api_source. The costs endpoint /organization/costs supports 1d only and groups by project_id, user_id, line_item, api_key_id and api_source; every amount carries a value and the currency usd.
Anthropic
The Usage and Cost Admin API needs admin credentials and is not available to individual accounts. /v1/organizations/usage_report/messages returns buckets of 1m, 1h or 1d (at most 1,440, 168 or 31 per request), split into uncached input, cached input, cache creation and output, and can be filtered and grouped by API key, workspace, model, service tier, context window, data residency (inference_geo) and speed. /v1/organizations/cost_report returns cost in USD as decimal values in cents, daily only, grouped by workspace or description. Code execution appears only in the cost endpoint, Priority Tier only in the usage endpoint. Anthropic states that data typically appears within five minutes.
The Gemini API runs on a Cloud Billing account, and its costs appear on the Cloud Billing cost management pages, usually within a day. Cloud Billing's export documentation names Google AI Studio among the products that report usage to Cloud Billing; the Gemini API's prepay balance is managed in AI Studio only. Gemini Enterprise Agent Platform (formerly Vertex AI) bills on Cloud SKUs, priced in USD or, for customers paying in another currency, in that currency on the Cloud Platform SKUs. Both therefore reach the Google Cloud billing export, with SKU, project and hourly window.
The hyperscalers' AI services
Amazon Bedrock appears on the AWS invoice and in the AWS Cost and Usage Report. Depending on the model, it bills by tokens, requests or images, or, for provisioned throughput, by hours per model unit. There are two allocation keys: tags on an application inference profile make the on-demand calls through it reportable as cost allocation tags, and with the caller identity option in a CUR 2.0 standard export, line_item_iam_principal records the calling IAM user or role for each Bedrock call, along with that principal's activated tags. A profile's price follows the model in the Region from which it is called.
In Microsoft Foundry (formerly Azure AI Foundry), Azure OpenAI and the other Foundry Tools run on Azure meters: standard deployments by input and output tokens, provisioned ones by hours per PTU regardless of use. They therefore sit in the Azure cost exports. In hub projects, models from Azure Marketplace appear per model offer and project as a resource of type microsoft.saas/resources, named [Model offer name]-[GUID], with meters such as paygo-inference-input-tokens and paygo-inference-output-tokens.
Claude bought through Amazon Bedrock or Google Cloud is invoiced by the cloud provider. Claude Platform on AWS and Claude in Microsoft Foundry bill through the respective marketplace in Claude Consumption Units (CCUs) at USD 0.01 each, reported hourly and shown as a single CCU line on the cloud invoice. For Claude Platform on AWS, Anthropic does not currently offer the Usage and Cost API, only the Usage and Cost pages of the Claude Console.
| Source | Interface | Units | Finest bucket | Allocation key | Invoiced by |
|---|---|---|---|---|---|
| OpenAI | usage endpoints, costs endpoint | tokens, requests, per-endpoint units; cost in USD | usage 1 minute, cost 1 day | project, user, API key, model | OpenAI |
| Anthropic | Usage and Cost Admin API | tokens by type; cost in USD cents | usage 1 minute, cost 1 day | workspace, API key, model | Anthropic |
| Gemini API, Gemini Enterprise Agent Platform | Cloud Billing, BigQuery export | per SKU | hourly window | project | Google Cloud |
| Amazon Bedrock | CUR 2.0 | per model: tokens, requests, images, model unit hours | as chosen for the export | account, inference profile tags, IAM principal | AWS |
| Microsoft Foundry | Azure cost export | token meters, PTU hours | day | resource, resource group, tags | Microsoft, partner models through Azure Marketplace |
What it means for your cost
A usage export is neither an invoice nor a price list. OpenAI and Anthropic deliver usage and daily cost in USD, but no invoice lines with an invoice number, payment or tax. How the providers bill is covered in AI provider billing, what they count in tokens, requests and context.
Usage is finer than cost: usage by the minute, cost by the day. An hourly cost curve is therefore derived, not a figure the provider supplies. For cost allocation, project, workspace, user (OpenAI) and API key are the finest keys the provider keeps itself; anything below that the application has to measure.
Usage bought directly sits beside the cloud exports in a source of its own and in USD. To appear in the same report it needs its own file import or API pull and a conversion into the reporting currency (CSV and file imports of cost data). Bought through a cloud, it is already in that cloud's export, but with the cloud's keys and, under CCU billing, without tokens or model. Neither provider's API reference states how long usage and cost data stay retrievable; a complete history exists only where it is pulled and stored regularly.
Where it goes wrong
- Cost is extrapolated from the usage endpoint. Symptom: the total differs from the cost endpoint. Cause: at Anthropic, code execution is missing from the usage endpoint and Priority Tier from the cost endpoint; the two do not cover the same things.
- A block without workspace or key. Symptom: part of the Anthropic usage cannot be assigned to a team. Cause: the default workspace appears as
null, and usage from the Claude Console playground has noapi_key_id. - Claude Platform on AWS is counted twice. Symptom: Claude cost appears twice in the report. Cause: the CCU line on the AWS invoice and the Usage and Cost pages of the Claude Console describe the same consumption.
- Bedrock cost has no team key. Symptom: Bedrock shows as one total per account and model. Cause: the calls do not go through an application inference profile with tags, and the caller identity option is not switched on in the CUR 2.0 export.
- Gemini API cost is missing from today's cloud report. Symptom: AI Studio shows consumption, Cloud Billing does not yet. Cause: usage reaches Cloud Billing with a delay, usually within a day, while AI Studio draws down the prepay balance continuously.
Sources
Read on 2026-10-16. OpenAI Usage, Completions and Costs; Anthropic Usage and Cost API and Pricing; Google Gemini API billing, Gemini Enterprise Agent Platform, Agent Platform pricing and Structure of Standard data export; AWS Set up a model invocation resource using inference profiles, Using IAM principal for cost allocation and Amazon Bedrock pricing; Microsoft Plan and manage costs for Microsoft Foundry hubs (classic) and Azure OpenAI pricing.