GitHubWhat they did
Sascha Heyer's team builds internal tools for departments across their company, and every pull request on those tools gets an AI review before a person reads it. The reviewer used to run on Claude, but the per-review cost added up across every push on every repo. They moved to gemini-review-action, an open-source GitHub Action maintained by another Google Developer Expert. It runs Gemini on Vertex AI, reads the PR with the wider codebase as context and posts inline findings on the pull request. They installed it on about 40 repositories with a single workflow file each. It signs in to Google Cloud through Workload Identity Federation, so no API key sits in any repo secret.
How it went
Between August and 14 October 2026 it ran 2,667 reviews for about $321 in total, roughly 12 cents each. On one early PR that a human had already approved with a bare LGTM, the reviewer found problems nobody had spotted. Running it at that scale exposed gaps, and Heyer sent the fixes upstream rather than forking: the bot now resolves its own threads once a finding is fixed, can ignore other review bots, reports cost per review, labels requests for billing, retries flaky GitHub calls and caps oversized files that used to blow the context.
Worth knowing
The cost figure only exists because every request carries a Vertex billing label, so add those first if you want to measure it. One huge generated file could exhaust the context and kill a review until the cap was added. Pin the action to a commit SHA and check concurrency settings, which could quietly run two reviews per push. The agent here is a third-party action calling Gemini models, not a Google product.