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.
Sources and attribution
OpenAI distinguishes granular usage evidence from the Costs endpoint and billing dashboard used for financial reconciliation.
AWS separates per-request invocation-log evidence from aggregated billed-dollar attribution in Cost Explorer and Cost and Usage Reports.
AWS recommends combining native billed-dollar attribution with per-request metadata when both invoice reconciliation and workload detail are required.