From the source
Anthropic published a guide for platform teams on controlling and measuring Claude Code spending on
Google Cloud, built around a control, track and prove framework.
- Claude on Google Cloud uses consumption-based pricing, so the bill reflects how many tokens tasks consume and which models handle them, and agentic work like Claude Code spends more tokens per task than chat because it reads files, runs tools and reasons across a whole task.
- Anthropic says high-cost sessions are often high value, usually power users on new workflows, while waste tends to sit with heavy spenders using a model like Fable for routine work Sonnet can handle.
- Google Cloud's native tools include per-project billing, budget alerts, spend caps in preview, BigQuery exports, per-minute quotas and provisioned throughput.
- The self-hosted Claude apps gateway runs in the customer's Google Cloud project, signs developers in through an identity provider so every request carries a verified identity, and sets model allowlists per IdP group plus spend caps per user, group or organization by day, week or month.
- Claude Code warns developers in the CLI at 75% and 95% of a cap, returns 429 at the cap until the period resets or an admin raises it, and caps can be changed through the gateway's admin API.
