Anthropic is rebroadcasting its session on controlling costs and showing return on investment for Claude Code on Google Cloud, this time for teams in Asia-Pacific time zones. The listing is short, but the topic is the one most engineering leaders get stuck on. Getting developers to try an AI coding agent is the easy part. Getting finance to approve the budget for another year is harder.
Anthropic’s listing doesn’t include a written walkthrough. What follows is a practical playbook built on how Claude Code and Google Cloud’s Vertex AI work together. You can use it before or after the session.
⚡ Quick Start
What you’ll learn: How to route Claude Code through your Google Cloud account, put limits on spend, track usage, and turn that data into an ROI story your leadership can check.
What you need:
- A Google Cloud project with Vertex AI enabled and access to Claude models
- Billing admin access, or a friendly colleague who has it
- Claude Code installed on developer machines
- A few weeks of real usage to measure
🛠️ The Playbook
1. Route Claude Code through Vertex AI
Set CLAUDE_CODE_USE_VERTEX=1, along with your CLOUD_ML_REGION and ANTHROPIC_VERTEX_PROJECT_ID.
Why it matters: Spend lands on your existing Google Cloud bill, under your committed-use agreements and IAM controls. Finance doesn’t have to track a separate vendor invoice.
2. Separate spend by team or project
Give each team its own Google Cloud project or clear billing labels.
Why it matters: A single “AI costs” line item can’t be defended. Spend broken out by team shows exactly where the money goes and who gets value from it.
3. Set quotas and budget alerts before rollout
Set Vertex AI quotas to cap throughput. Then create budget alerts at 50%, 90% and 100% of your monthly target.
Why it matters: An agent running in a loop, or one power user, can burn through a month’s budget fast. Guardrails you set on day one cost less than an awkward conversation on day thirty.
4. Match the model to the task
Use the most capable model for hard reasoning and architecture work. Send routine edits and background tasks to faster, cheaper models.
Why it matters: Model choice is the biggest cost lever you have. Most coding work doesn’t need the top-tier model.
5. Keep prompt caching on
Claude Code reuses large chunks of context across turns, and caching makes those repeated tokens much cheaper.
Tip: Don’t turn off caching to fix unrelated issues. It’s one of the quietest ways to save money.
6. Turn on telemetry
Set CLAUDE_CODE_ENABLE_TELEMETRY=1 and export OpenTelemetry metrics to your monitoring stack. Claude Code reports sessions, token usage, cost, lines of code changed, commits and pull requests.
Why it matters: You can’t prove ROI from anecdotes. This is your raw data.
7. Export billing to BigQuery
Turn on Cloud Billing export and join it with your telemetry.
Why it matters: Now you can answer “cost per developer” and “cost per merged PR” with a query, not a guess.
8. Build the ROI case on outcomes
Compare cycle time, PR throughput and time spent on chores like tests, migrations and docs before and after adoption. Set those gains against cost per developer.
Warning: Lines of code is a weak metric on its own. Leadership will trust delivery speed and fewer hours spent on grunt work much more.
📊 Why This Matters
What stands out is that Anthropic is putting cost governance front and center. It’s not only selling capability. Enterprise AI budgets are moving out of experiment funds and into operating plans, and those plans need clear numbers. If a team can’t show cost per outcome, it’s likely to lose its seats in the next budget cycle.
Routing through Google Cloud also matters for regulated industries and APAC companies with data residency rules. Picking a region and keeping existing IAM controls often decides whether legal approves a rollout at all.
🚀 Next Steps
- Run a 30-day pilot with one or two teams, with telemetry on from the first day
- Set a baseline now for cycle time and PR volume, before adoption changes the numbers
- Share a monthly one-page report with cost, usage and outcomes so finance sees the trend
- Review model routing every quarter as new, cheaper models come out
- Put org-wide settings in managed configuration so individual developers can’t quietly override cost controls
As AI coding agents take on longer autonomous tasks, token use per developer will keep rising. The teams that measure it now will be the ones that get more budget later. You can find the full session details in Anthropic’s original listing.