Billing / AI Foundry metrics do not match

Question

Billing / AI Foundry metrics do not match

Raymond Pitts 0

Azure AI Foundry / Azure OpenAI Monitoring Reconciliation Bug

We are reporting a monitoring / billing reconciliation failure in Azure AI Foundry / Azure OpenAI that materially affected our ability to estimate and control cost.

Impact

We relied on Azure monitoring surfaces to understand token consumption and expected spend across our AI Foundry deployments. The numbers exposed by Azure did not reconcile with our real-time usage experience or with Azure billing exports, and our actual cost was materially higher than expected.

This was not a small rounding issue. For several GPT-family deployments, the discrepancy is large enough that it undermines confidence in Azure’s monitoring for operational cost control.

Problem statement

For the same subscription and overlapping May 2026 usage window, we observed a mismatch between:

real-time operational experience using our Azure AI Foundry deployments,
Azure Monitor deployment-level usage metrics, and
Azure billing CSV exports.

These three views do not reconcile.

The discrepancy is especially visible on GPT-family models. Some simpler categories, such as embeddings and Kimi input, align much more closely, which suggests the issue is concentrated in GPT-family monitoring / billing attribution rather than our extraction logic.

Subscription / scope

Subscription ID: `

Raymond Pitts 0 Reputation points

2026-06-02T01:59:46.7633333+00:00

The model aliases showing Opus/Sonnet/Haiku are just the aliases
Bharath Y P 10,530 Reputation points Microsoft External Staff Moderator

2026-06-02T02:48:18.0766667+00:00

Hello Raymond, thank you for posting your query on Microsoft Q&A platform.

We are looking into your issue and will keep you posted updates.

Thanks
Bharath Y P 10,530 Reputation points Microsoft External Staff Moderator

2026-06-02T04:52:51.25+00:00
Hello Raymond, You are reporting a cross-surface reconciliation failure between Azure Monitor (deployment-level token metrics), Azure AI Foundry operational experience (runtime usage) and Azure Cost Management / billing CSV exports (metered charges). Impact:

GPT-family deployments show material mismatch across input/output tokens

This prevents real-time cost estimation and control, leading to unexpected higher spendYou are reporting a cross-surface reconciliation failure between:

Azure Monitor (deployment-level token metrics)

Azure AI Foundry operational experience (runtime usage)

Azure Cost Management / billing CSV exports (metered charges)

Impact:

GPT-family deployments show material mismatch across input/output tokens

This prevents real-time cost estimation and control, leading to unexpected higher spend

1. Azure Monitor vs. Cost Management surface - Monitor metrics (inputTokens, outputTokens, totalTokens) are telemetry-oriented and may include aggregation, latency, or batching. - Cost Management data (billing CSV / Cost Analysis) represents the authoritative metered usage used for invoicing and is typically available with a delay (commonly up to ~24 hours).

2. GPT-family attribution nuances - Azure OpenAI usage is billed at the meter/SKU level rather than strictly at the deployment level. - As a result, deployment-level monitoring metrics may not map 1:1 to billing categories, particularly for GPT-family models, whereas simpler models (like embeddings) tend to align more directly.

3. Why rounding or latency isn’t the culprit - The observed discrepancies are orders of magnitude larger than typical rounding or ingestion delays. - The closer alignment for simpler models (embeddings, Kimi) suggests the issue is concentrated in GPT-family attribution or aggregation.

4. What you can do for cost visibility - Enable Cost Management exports to obtain detailed meter-level usage data. - Use Cost Analysis to group by Meter or Service and analyze GPT-related billing meters. - Treat Cost Management as the authoritative source for cost, and Azure Monitor as an operational telemetry view.

Azure Monitor deployment metrics and Azure billing exports can diverge: Azure Monitor provides near real-time telemetry metrics that are aggregated and may include ingestion latency or batching. - Azure billing exports are generated from the metering system and represent authoritative, invoice-based consumption. - Because these systems use different aggregation pipelines, categorization, and data processing stages, exact reconciliation is not guaranteed.

Share authoritative documentation on billing surfaces and APIs - Azure Cost Management (Cost Analysis, Usage Details, exports, and Consumption APIs) is the authoritative source for billing data. - However, publicly documented mappings between ServiceResource categories and Azure OpenAI deployment-level metrics are limited.

Monitor metrics are expected to match billing exports - Azure Monitor metrics are not expected to exactly match billing data. - Monitor metrics serve operational and diagnostic purposes, while billing exports reflect metered usage used for invoicing.

Preferred surface for real-time cost control and forecasting - Azure Cost Management (Usage Details API, exports, Cost Analysis) should be treated as the authoritative source for cost tracking. - API-level usage data (e.g., OpenAI usage responses) can support near real-time estimation but should not be relied upon for exact billing reconciliation.

Correlate deployment usage with billing by grouping costs by Meter, Service, or Resource in Cost Analysis. - Normalize units (e.g., per 1K or 1M tokens) and align time windows before comparison. - If discrepancies persist after reconciliation, open a support request to investigate backend metering and attribution.

Reference docs

Troubleshoot inconsistent or lagging cost data

Compare billed charges with usage file & Cost Analysis

Plan and manage costs for Microsoft Foundry (meter mapping, grouping):

Fine-tuning & hosting charges guidance

1 answer

Your answer

Raymond Pitts 0 Reputation points

2026-06-02T01:59:46.7633333+00:00

The model aliases showing Opus/Sonnet/Haiku are just the aliases
Bharath Y P 10,530 Reputation points Microsoft External Staff Moderator

2026-06-02T02:48:18.0766667+00:00

Hello Raymond, thank you for posting your query on Microsoft Q&A platform.

We are looking into your issue and will keep you posted updates.

Thanks

Answer 1

Raymond Pitts 0

@Bharath Y P That is normalized for the time window and token counts (1M vs 1K), and is not from the responses but from the AI Foundry API itself.

We have only a few deployments per model and each was aggregated for those numbers, and they still are way off the mark.

We attempted to create a ticket, but they have not responded yet.

Please advise.

Raymond Pitts 0 Reputation points

2026-06-02T21:15:40.17+00:00

Correction: Azure Portal did not give us the option to create a support ticket, only to post here.
Raymond Pitts 0 Reputation points

2026-06-03T01:49:04.3966667+00:00

The account we are working from is a sponsorsed Azure for Startups using credits.. When we click "request a refund" it says:

Your account is not eligible for expedited refunds

Submit a support request to get a refund.

and then takes us to the same page.
Bharath Y P 10,530 Reputation points Microsoft External Staff Moderator

2026-06-03T02:01:32.3433333+00:00
Hello Raymond Pitts, Thanks for the response, you can follow the below steps to request for the refund manually.

Here are the steps you can follow to create support requests:

Go to Azure portal > Search for Help + Support or https://portal.azure.com/#view/Microsoft_Azure_Support/HelpAndSupportBlade/~/overview

Click on Create a Support Request.

On right-side panel, enter Billing and click Go > select the issue with "Billing" > select your Subscription then click Next.

Do not click on refund request button or any suggested articles instead click on the Create support request on top of the panel.

Select the Issue type as Billing and, select the summary, problem type and problem subtype as per your issue like below > then click on Next.

Do not click on refund request button or any recommendation solution/article, instead click on the Return to support request then click Next.

Fill all the required information and provide a clear description of your issue and click Next.

Review and create as submit the support request.

You should receive an automated email within a few minutes and then be contacted by support team.

Kindly check the below documentation for more details: How to create an Azure support request - Azure supportability | Microsoft Learn

Hope this helps! If you encounter any issues, please reach out to us. Thanks.
Bharath Y P 10,530 Reputation points Microsoft External Staff Moderator

2026-06-05T08:36:17.6233333+00:00

Just checking in to see if the solution shared above help you to resolve your issue. please reach out to us If you have any further questions.

Thanks
Raymond Pitts 0 Reputation points

2026-06-05T10:53:57.7466667+00:00

@Bharath Y P it did not. We have had zero support / response via our Azure portal support request.
Raymond Pitts 0 Reputation points

2026-06-07T17:08:10.93+00:00

@Bharath Y P please advise on what my team is to do regarding unanswered support requests and the status on how the issue is being investigated.
Adriaan 85 Reputation points

2026-07-15T15:48:12.79+00:00

Hi @Raymond Pitts ,

I just want to say I feel you. We have been disputing a similar ~$12000 overcharge on supposed longco (long context) tokens that we are unable to reconcile to any metrics in Azure or that we have on hand. And this has been going on since the 5th of June, after noticing our sponsored credits had been wiped out at the beginning of May!

We ended up having to buy a support plan just to get someone to respond, and even then, we are no closer to a response or clarification on the issue. We've had to fight our way through multiple responses of "the billing matches the metering" nonsense, without them attempting to even understand that even though that appears to be the case, the metering is definitely wrong.

Anyway, have you had any further luck?