Google Vertex AI and Gemini

Connection steps are in the Vertex AI and Gemini connector guides. This page is only attribution. Both connectors read the same kind of billing export, with a different service filter.

What you set up

1. Turn on the Cloud Billing export to BigQuery for the billing account. Without that export, Vertex AI has nothing to pull. Gemini still validates a key, and it does not pull spend until the export is there.
2. Connect the provider with its connector guide, using credentials that can query that export.
3. Create departments and agents you will assign to, under Governance and Agents & allocation.

Google can copy request labels onto Cloud Billing for pay-as-you-go calls. (Add labels to API calls) The recommended keys are on the tagging page. Meridian's pull does not group by those labels, or by project id beyond the billing account you connected. Labels you send are not what this connector attributes.

A Cloud project is the billing boundary inside a billing account. One project shared by many agents is one set of daily SKU rows.

What Meridian receives

One charge per UTC day and SKU. The query groups by day, SKU, and currency. Labels, resource names, and the project id on each export row are not stored as the agent or the department.

The export's own time windows are often hourly, and many calls can share a row before Meridian sums them to a day and a SKU. (Standard usage export) Costs can take more than a day to appear.

Gemini without a billing export does not write spend rows. The key check can still succeed.

How it's attributed

Rows arrive unattributed. A rule can match the vendor or the model. A person can assign the inbox rows. The pull does not declare an agent and does not inherit a label.

A rule fills only rows still unattributed. It does not replace a mapped, inherited, or declared row from another path, and it does not replace a row a person has assigned.

What lands unattributed, and how to fix it

Every day-and-SKU total from this pull lands unattributed until you assign it or a rule matches it. The Reason column has no label sentence, because labels were not read.

Assign the rows, or add a rule and choose Apply rules now. Unlocked dates update. A locked month shows the change as an adjustment later.

To separate agents, send labels only if you also have another path that can see them. This pull cannot. Use a gateway or OASA events per agent, or a FOCUS upload if your Google export is in FOCUS and the labels survived into that file. Which BigQuery column holds a generateContent request label is not settled in Google's export schema. Meridian does not guess it.

Limits

• Provisioned Throughput silently ignores request labels. The call still succeeds. The label is not on the bill. (Add labels)
• The labels page's examples use Vertex (aiplatform.googleapis.com). A labels-to-billing page for the Gemini Developer API (generativelanguage.googleapis.com) was not part of the provider pages this guidance relies on. Do not assume the same labels body reaches a Gemini Developer API bill.
• A label key can be dropped after 1,000 distinct values for that key over the life of the billing account. That is Google's rule. It does not change this pull, which does not read labels either way.
• Vertex and Gemini are two catalog entries and two connections. They share this attribution behavior. A dollar that is in one export filter is not in the other.