Cloudflare Cut the Cost of Each AI Code Review by 68 Percent

Cloudflare published a detailed account on 19 October 2026 of how it cut the cost of its AI code reviewer. Average cost per review fell from about $2.43 to $0.77, a 68 percent reduction, with the code quality reviewer still set to high effort. Cost had roughly doubled from about $1.19 in April to about $2.46 by late August before the work began.

Speed improved as well. Median review time dropped from 4m 09s to 1m 32s, and the 99th percentile from 15m 20s to 4m 41s. About 40 percent of the savings came from code changes to the reviewer, with the rest from testing different models and effort levels.

The most transferable part is prompt caching. Cloudflare moved the shared prefix, the tool definitions and agent prompt of roughly 11k to 21k tokens, into its own cache breakpoint, which lifted the first-request cache hit rate from 14 to 26 percent up to 91 to 99.5 percent. It also made its three sub-reviewers (code quality, security and performance) share an identical first request, so one writes the cache and the other two read it.

On models, the coordinator moved to Opus 5.5 at medium effort, and the Sonnet reviewers moved to Sonnet 5.5, which Cloudflare measured at about 47 to 62 percent cheaper per run than Sonnet 5 at the same per-token price. Authors acted on 59 percent of the code quality reviewer's serious findings, up from 34 percent.

Cloudflare is honest about the trade-off. Performance review at medium effort appears to have lost real findings, and security may have too. Testing both at high effort is the next step.

The lesson for anyone running agents in a pipeline is to measure cost per task, put the stable part of a prompt first so it can be cached, and check quality every time you lower effort. Cheap and wrong is not a saving, and Cloudflare's willingness to say so is what makes the post worth reading.