Cloudflare reports 68% lower AI code-review costs—with quality trade-offs

Cloudflare reports cutting the average cost of its internal AI code reviewer to $0.77, down 68% from the $2.43 average during its two most expensive weeks before optimization. The cost figures assume high effort for the code-quality reviewer—a setting Cloudflare says it is moving to after evaluating quality trade-offs. Median review time fell from 4 minutes 9 seconds to 1 minute 32 seconds. Cloudflare’s report

Cloudflare attributes roughly 40% of the savings to changes in the reviewer’s code and the rest to model and effort choices. It cautions that its release-by-release cost attribution was not a controlled experiment. Its quality evaluation also found different outcomes across specialist reviewers: lower costs did not consistently preserve the number of important findings. Cloudflare

A high cache hit rate can hide growing costs

The reviewer’s cache hit rate rose from 85.7% in April to 97% by September, while tokens per review increased from 0.91 million to 2.99 million. For its code-quality reviewer, Cloudflare says each model call sent the whole conversation as context. Repeated calls over growing histories therefore increased spending even when most tokens qualified for discounted cache reads. Cloudflare

Cloudflare separated reusable prompt content from repository-specific details so different jobs could share cached prefixes. It also removed tools that sub-reviewers were not permitted to use, arranged parallel file reads and eliminated a final model turn that reread the conversation to produce an unseen summary. Removing that final turn saved about $0.07 per job, according to the company. Cloudflare

The practical lesson is to track context volume and call count alongside cache hit rate. Discounted repetition is still spending; a high hit rate alone does not establish that an agent is working efficiently.

Quality depended on the reviewer and effort setting

Cloudflare concluded that its code-quality reviewer at high effort continued to find roughly as many important problems as before, with much of the reduction in critical and warning findings attributed to noise and inflated severity. In its randomized comparison, however, code quality at medium effort produced roughly a quarter as many acted-on findings as high effort. That result prompted the move to high effort. Cloudflare

The performance reviewer showed a different trade-off. Cloudflare’s model judges counted about 64% fewer important performance findings per review—equivalent to about one fewer important finding every 24 full-tier reviews. Important security findings fell by about a quarter, although Cloudflare says that result was not statistically significant. Both reviewers remained at medium effort, with high-effort testing planned next. Cloudflare

These evaluations have limits. Cloudflare counted a finding as acted on when the author’s next push changed the flagged lines; that does not establish that the finding was correct. Its separate assessment used models to judge a blinded sample of 657 findings, with only 35 to 66 findings in each comparison group. The report also says its “now” comparison window was only one day long. The results are useful evidence about this deployment, not a definitive measure of defect-detection accuracy. Cloudflare

Separately, Cloudflare’s OpenCode v1/v2 shadow comparison found 98% agreement on approval decisions in canary repositories. That measures agreement between the two implementations, not whether either missed defects, and it does not establish that every cost optimization preserved quality. Cloudflare

For engineering teams, the useful distinction is between eliminating wasted work and reducing review effort. Removing an unused summary turn is a different intervention from changing how intensively a specialist searches for problems. Evaluate those changes separately rather than treating the total cost reduction as a single quality-neutral improvement.

Internal rollout and experimental infrastructure

Cloudflare says it moved all internal reviews to OpenCode v2 in early October. Running the reviewer solely on Durable Objects using pi-durable was still experimental in the report. Those are distinct stages: a completed internal migration and an infrastructure experiment. Cloudflare