Claude models on Vertex AI are partner models — fully managed, serverless models enabled through Model Garden and billed against the Google Cloud project that calls them. There is no separate Anthropic invoice: token charges appear alongside your other Google Cloud spend, priced from the "Generative AI on Vertex AI" pricing page under "Anthropic's Claude models." That makes cost management familiar territory for anyone who already runs GCP FinOps, with a handful of Claude-specific rates worth memorizing.
The rates that show up on the bill
Google's published global-endpoint list prices per million tokens match Anthropic's first-party rates: Claude Opus 4.8 at $5 input / $25 output, Claude Sonnet 5 at $3/$15 standard (promotional $2/$10 through August 31, 2026), Claude Haiku 4.5 at $1/$5, and Claude Fable 5 at $10/$50. Around those base rates, four modifiers can appear:
| Modifier | Effect |
|---|---|
| Prompt caching | 5-minute cache write 1.25x input; 1-hour write 2x; cache hit 0.1x (Opus 4.8: $6.25 / $10.00 / $0.50 per MTok) |
| Batch prediction | 50% of on-demand ("Batch Input"/"Batch Output" rates, e.g. Opus 4.8 batch $2.50/$12.50) |
| Regional / multi-region endpoints | 10% premium over global (e.g. Opus 4.8 on us: $5.50/$27.50), for Sonnet 4.5 and later |
| Long context | Google's pricing page states that when input context is ≥ 200K tokens, all tokens bill at the per-model long-context rates — check the >200K columns for your model |
Web search tool usage bills separately at $10 per 1,000 searches. Provisioned Throughput, if you buy it, is a fixed-fee subscription arranged with sales and available only on regional endpoints — a different line entirely from pay-as-you-go tokens.
Billing export to BigQuery
For anything beyond eyeballing the console, the standard mechanism is Cloud Billing export to BigQuery: it lands every charge as a row you can query, joined with SKU, project, and label dimensions. The exact SKU descriptions for Claude usage encode model and token direction, and they evolve as Google adds models and endpoint types — so rather than copying SKU strings from a blog post, run a query grouped by SKU description over a day of real traffic and build your filters from what your own export contains. Google's pricing page is the authoritative mapping from SKU to rate; where the two seem to disagree, check the official documentation before escalating.
A practical starting query pattern: filter to your Claude-serving projects, group by SKU description and project, and sum cost by day. That alone gives model-level trend lines suitable for a weekly FinOps review.
Label-based attribution and its limits
Google Cloud's cost attribution model leans on projects first and labels second. For Claude workloads the honest guidance is:
Projects are the reliable boundary. Each Claude-calling project accrues its own charges, so putting each team, product, or environment in its own project makes the billing export attribute spend with zero extra work. This mirrors the general 3P pattern described in separating dev, staging, and production costs.
Labels work where resources exist to label. Serverless partner-model inference doesn't give you a long-lived resource per application the way a VM does, so label coverage for Claude token spend depends on how your organization structures projects and billing metadata. If a slicing dimension you need isn't represented, capture it yourself at the application layer — log the usage object from every response with your own team/feature metadata, as described in logging token counts for internal FinOps.
Cross-checking tokens against the bill
Vertex AI does not offer Anthropic's Usage and Cost Admin API, so your reconciliation dataset is Google-side. Two caveats from Google's own documentation matter here. First, token usage shown on the console Quotas page may be inaccurate due to Anthropic's token estimation and refund system — use the token counting API or the token_count metrics in Metrics Explorer for accurate usage instead. Second, if you enable request-response logging to build your own per-request token dataset, note that it isn't available together with VPC Service Controls. With clean token logs in hand, multiplying by the pricing-page rates should land close to the invoice; the methodology for explaining residual gaps is covered in reconciling cloud invoices against internal token logs.
Where to go next
Compare the other clouds' mechanics in AWS billing for Bedrock and Azure Cost Management for Foundry, or start building attribution with tagging Claude requests.