Anthropic outlines cost controls for Claude Code on Google Cloud

Anthropic has published guidance for organizations using Claude Code on Google Cloud, arguing that rising adoption needs to be matched by tighter cost controls, better usage tracking and a clearer way to attribute business value. The company frames the approach as three levers — control, track and prove — in a post published Oct. 27, 2026.

Cost controls center on tokens, models and gateway limits

According to Anthropic, Claude Code runs on consumption-based pricing, so monthly spend reflects how much developers use the tool in a Google Cloud environment. The company says token consumption rises with task complexity and conversation length, and that agentic work through Claude Code generally uses more tokens than a basic chat interaction because it reads files, runs tools and reasons across a whole task.

Anthropic says organizations can influence spend through Google Cloud-native controls such as per-project billing, budgets and alerts, spend caps in preview, BigQuery exports, quotas and provisioned throughput. It also points to Claude apps gateway controls that can set model allowlists and spend caps per user, identity-provider group or organization, with limits adjustable by day, week or month. The gateway is described as a self-hosted service shipped in the same Claude binary developers already use, and requests are tied to verified identities through an identity provider such as Google Workspace.

Tracking ties usage data to cloud observability

The company says Claude Code can export OpenTelemetry metrics and events covering usage and estimated cost, with breakdowns by skill, plugin, agent and MCP server. Anthropic says those metrics can be sent to monitoring tools such as Google Cloud Monitoring through an OpenTelemetry collector or via the Claude apps gateway, and that Cloud Billing exports to BigQuery can be used to analyze Claude SKUs alongside other cloud costs.

Anthropic also says the gateway can apply telemetry settings centrally, tag exports with verified email and identity-provider groups, and add custom attributes from the OTEL_RESOURCE_ATTRIBUTES environment variable so teams can aggregate by department or cost center. The post says a ready-made Cloud Monitoring dashboard template and a quickstart are available for setting up telemetry, logging and BigQuery tables for API requests, tool decisions, tool results, assistant responses and completed hook executions.

ROI is measured through usage, satisfaction and impact

Anthropic says the final step is proving return on investment by measuring usage, developer satisfaction and impact. It says adoption metrics can show activated users, daily and weekly usage, retention and power-user behavior; satisfaction metrics can be collected through surveys and check-ins; and impact can be assessed by comparing throughput, cycle time and review speed against an internal baseline.

For impact, the post suggests tracking time, effort and volume metrics, including PR cycle time, code review turnaround, review iterations per pull request, review comments, deployment frequency and named bodies of work shipped with freed-up capacity. Anthropic says the baseline already exists in source control and ticketing history and recommends defining a consistent unit of work, since not every task is a pull request.

What Anthropic says customers are seeing

Anthropic cites examples from several customers to illustrate the framework. It says Delivery Hero used Claude Code to overhaul a design system across 100 UI components in minutes rather than about a week of an engineer's time. It says Shopify uses the tool to help engineers build faster and to let non-engineers build internal tools in minutes instead of days. It also says Fifth Dimension’s engineers are using Claude Code on Google Cloud broadly and are writing more than 300,000 lines of code a month with a 98% acceptance rate.

The company’s guidance is aimed at platform teams trying to balance access with governance as usage grows. Anthropic says the process should start with default controls such as an org-wide spend cap, a default model and a model allowlist, then be refined by role and team as telemetry and cost data come in. The post does not give a date for a future product milestone or event beyond the publication date of the guidance.