One of the quiet advantages of running Claude through Vertex AI is that there is no separate Anthropic invoice: consumption is metered by Google and appears in the GCP Billing console alongside your compute and storage. Billing is pay-as-you-go per million tokens (or a fixed fee if you've purchased provisioned throughput), and the authoritative rate card is Google's "Generative AI on Vertex AI" pricing page, under the "Anthropic's Claude models" section. This article maps what you'll see in the console back to that rate card.
Finding the Claude line items
In the Billing console's cost table and reports, filter to the Vertex AI service and then group by SKU — Google's term for an individually priced billing unit. Google prices Claude by model and by consumption category, so expect a family of SKUs per model rather than one. The categories that exist per the pricing page:
| Rate category | Rate (vs on-demand input/output) | Opus 4.8 example (per 1M tokens) |
|---|---|---|
| Input / output tokens (global endpoint) | List price | $5.00 / $25.00 |
Input / output on us or eu multi-region and regional endpoints | +10% over global | $5.50 / $27.50 |
| Batch input / batch output (Google-native batch) | 50% of on-demand | $2.50 / $12.50 |
| Cache write (5-minute / 1-hour) | 1.25x / 2x input | $6.25 / $10.00 |
| Cache hit | 0.1x input | $0.50 |
| Web search | Per-use | $10 per 1,000 searches |
Exact SKU description strings vary and are best confirmed against the pricing page itself; the reliable pattern is that each combines a model name with one of the categories above, so grouping the cost table by SKU effectively gives you cost by model and consumption type.
Three multipliers that explain surprise variances
Endpoint premium. Regional and multi-region endpoints cost 10% more than the global endpoint for Claude Sonnet 4.5 and later models. If a team quietly pinned to us-east5 for latency, their per-token cost rose 10% relative to your global-endpoint estimate. The endpoint decision guide covers when that premium is worth paying.
Long-context surcharge. Google's pricing page states that if a query's input context is 200K tokens or longer, all tokens in the request — input and output — are charged at long-context rates (see the per-model >200K columns on the pricing page). A workload that occasionally stuffs the 1M-token window will show noticeably higher effective per-token rates than one that stays under the threshold.
Promotional pricing with an expiry. Claude Sonnet 5 bills at a promotional $2 input / $10 output through August 31, 2026, then $3 / $15 from September 1. If your forecast was built on the promo rate, expect a step change in the September invoice.
Setting up a recurring cost view
For a dashboard you'll revisit monthly, build a saved cost-table view filtered to the Vertex AI service and grouped by SKU and project — that yields cost by model, category, and consuming team in one screen (assuming teams are separated by project, the cleanest attribution unit on GCP). For finer-grained slicing — per feature or per environment within one project — export billing data to BigQuery and query it, as covered in the cost attribution article.
One reconciliation caveat from Google's own documentation: token usage shown on the console Quotas page may be inaccurate due to Anthropic's token estimation and refund system. For usage numbers you intend to reconcile against the bill, use the token counting API or the token_count metrics in Metrics Explorer instead.
If you've purchased provisioned throughput — Google's fixed-cost, fixed-term subscription that reserves Claude capacity on regional endpoints — that appears as its own fixed fee rather than per-token consumption, and the reading exercise changes accordingly; see the provisioned throughput guide.
Where to go next
Pair this with cost alerts and budgets so surprises page you before the invoice does, and the estimation worksheet for forecasting new workloads.