Cloudflare's October 19 engineering post describes an internal effort to reduce the cost of its AI code reviewer. The company reports a 68 percent drop in average cost per review from its most expensive weeks, alongside a shorter median review time.
Where the savings came from
- Cloudflare attributes about 40 percent of the savings to changes in the reviewer's own code.
- The remaining work included experiments with model selection and how much effort each model used.
- The post also describes a higher prompt-cache hit rate, while noting that total tokens per review had grown.
These measurements describe Cloudflare's own workload and do not predict the costs of another codebase. The post emphasizes measuring token use, model quality, latency, and review tiers together rather than optimizing one number in isolation.
For teams building AI-assisted reviews, compare representative changes before and after each optimization and keep human review responsible for important decisions.