Learning center

LLM cost attribution: reconcile provider bills

Compare provider financial totals with request-level gateway evidence, investigate variance, and allocate cost without uploading the source bill.

For FinOps, finance administrators, platform owners, and portfolio-company operators · Updated September 1, 2026

1. Choose an exact comparison scope

Open Reconciliation in the customer console. Choose the provider and, if the statement applies to only one gateway provider account, choose that account. Use UTC dates and treat the end as exclusive: a July statement runs from July 1 through August 1.

Either enter the USD provider subtotal and an optional internal label, or choose the normalized CSV template. The browser calculates the subtotal locally and submits only the summary and basic file counts. The raw file, filename, line items, and local notes are not uploaded.

The saved summary records what the operator submitted for comparison. It does not authenticate provider origin, prove completeness, validate negotiated rates, establish settlement, or provide accounting approval.

Why request-level cost and the provider bill can differ

Request metadata answers which workload, route, model, and principal used AI. A provider financial source answers what the provider actually billed for the period. Contract discounts, credits, taxes, service tiers, caching, rounding, delayed usage, and activity outside the gateway can make those totals differ.

Do not force either source to impersonate the other. OpenAI distinguishes granular usage from its financial Costs endpoint. Amazon Bedrock distinguishes per-request token evidence in invocation logs from billed-dollar attribution in Cost Explorer and Cost and Usage Reports. A defensible allocation keeps both sources, their time windows, and the remaining variance visible.

2. Read the evidence status before the variance

The comparison uses at most the latest 1,000 retained gateway request-metadata events. It reports matching requests, provider-reported versus modeled cost, an open period, and a truncated evidence window. Those limitations determine whether the result can be within tolerance or must remain incomplete.

A positive variance means the provider subtotal is above gateway-observed provider cost. A negative variance means gateway-observed cost is above the entered subtotal. No result proves invoice completeness, negotiated-rate accuracy, payment status, or accounting approval.

3. Allocate the subtotal for showback

When matching request evidence exists, the console proportionally allocates the exact entered subtotal across workload, environment, client type, data class, route, and model. Every dimension sums back to the subtotal, with small categories bounded into Other.

Use the allocation to investigate ownership and showback. It is decision support, not a ledger posting or chargeback approval.

4. Correct evidence without erasing history

Statement summaries are immutable. If the provider, scope, period, or subtotal is wrong, void the summary with a reason and enter a replacement. The void and its audit history remain available, while only active summaries contribute to portfolio finance reporting.

5. Share only the portfolio evidence a company approves

A portfolio sponsor first requests aggregate reporting. Statement subtotal and variance require the separate finance-reconciliation scope in the same explicit request, and the company owner sees both scopes before deciding. Without that finance scope, portfolio screens, CSV, and board PDF contain no statement figures.

Every portfolio read revalidates the company link. Revocation removes statement totals and variance immediately. Raw invoice files, labels, provider-account identifiers, request-level detail, payment data, and credentials are never included in the portfolio rollup.

Sources and attribution

OpenAI API — Usage

OpenAI distinguishes granular usage evidence from the Costs endpoint and billing dashboard used for financial reconciliation.

Amazon Bedrock — Track usage and costs

AWS separates per-request invocation-log evidence from aggregated billed-dollar attribution in Cost Explorer and Cost and Usage Reports.

Amazon Bedrock — Cost attribution best practices

AWS recommends combining native billed-dollar attribution with per-request metadata when both invoice reconciliation and workload detail are required.

Put it into practice

See the working controls behind the guide.

Run a bounded provider simulation without a login, card, key, or model charge. Create a workspace only when you are ready to save controls or connect live traffic.

Create a free workspace Try the Test LabTalk with us