Cloudflare cuts AI code review costs 68% through optimization

Cloudflare's average AI code review cost had risen from $1.19 in April to about $2.46 in August, prompting the team to aggressively cut token usage starting mid-September. After optimization, a review at high reasoning effort costs just $0.77 on average, down 68% from $2.43 in the prior two weeks, and median review time fell from 4m09s to 1m32s. About 40% of the savings came from reviewer code changes, the rest from model selection and reasoning-effort adjustments. Specific changes: reusing a shared system-prompt prefix across repos on each session's first request (cache hit rate was previously only 14%–26%); moving the internal standards index into the prompt, cutting that reviewer's calls from 15.9 to 9.6 and per-call cost from $1.12 to $0.72; removing tools sub-reviewers don't use, reading files in parallel, skipping reviewers with no new content on re-review, stopping immediately after posting comments, and periodically checking whether a task was cancelled. On the model side, the orchestrator moved uniformly to Opus 5.5 at medium reasoning effort; on October 3 the Sonnet reviewer was split 50/50 between Sonnet 5 and 5.5, and after two and a half days 5.5 was 62%, 58%, and 47% cheaper in code quality, performance, and security scenarios respectively with comparable key findings, leading to a full switch on October 5. Teams maintaining multi-agent review pipelines can learn from this: first quantify where tokens go, then adjust cache prefixes and call counts.