Cloudflare cuts AI code review costs by 68% to $0.77 via prompt caching and model optimization

Details

Average review cost dropped from $2.43 to $0.77 (68% reduction) and median latency fell from 4m 09s to 1m 32s after aggressive token optimization starting in mid-September.
40% of savings came from code changes, primarily fixing prompt cache breakpoints so shared system prompts are read from cache instead of written, reducing cache writes by 54% and reads by 82%.
Model and effort adjustments contributed the remaining 60% of savings; switching to Sonnet 5.5 at medium effort cut LLM calls per code quality review from 6.2 to 1.7, though high effort was retained for code quality to preserve finding quality.
Review quality held up for code quality at high effort, with acted-on findings per 1,000 reviews remaining stable (132 vs 145), while severity inflation dropped 70% as the reviewer stopped flagging minor issues as critical.
Reliability improved with reviews ending without a verdict dropping from 0.79% to 0.27% and security reviewer retries falling from 3.5% to 0.04%.

Key point