CLOUDFLARE’S AI code reviewer has been transformed from a cost-heavy system to a leaner, faster workflow. After a period of rising costs, the team focused on where tokens were being spent and implemented aggressive optimisations across the review process. The result: the average cost per review fell from about $2.43 to $0.77, while the median review time shrank from 4 minutes 9 seconds to 1 minute 32 seconds.
Roughly 40% of the savings came from code-level changes, with the rest from tweaking which models were used and how much effort they exerted. The project also delivered a notable efficiency boost in reviewer throughput.
Key technical shifts included moving reviewers onto Anthropic models (which raised per-job costs initially) and upgrading to OpenCode v2. A major win came from reworking the prompt caching strategy: sharing a cached prefix across repositories dramatically reduced token use and calls. The cache hit rate climbed from 85.7% in April to 97% by September, and tokens per review more than tripled, yet costs fell as the system wrote and read fewer tokens overall.
The team also consolidated first-requests across three reviewers to reuse identical prompts and tools, while keeping reviewer-specific instructions separate. By early October, the combined effect of prompt sharing, model/tier adjustments, and token management delivered the large cost and latency reductions documented in the post.