Cloudflare cuts AI code review costs by 68%

💻 Cloudflare cuts AI code review costs by 68%

Cloudflare optimized its AI code reviewer: the cost of a review dropped from $2.43 to $0.77, and the median time from 4:09 to 1:32. In August, the bill increased after switching to Anthropic: tokens per review went from 0.91M to 2.99M.

🌍 A rare public breakdown of the unit economics of an agentic pipeline: the leak was not in the pricing. The first request of a session was writing to the cache instead of reading (hit rates of 14–26%) — OpenCode v1 mixed the prompt with repository data. After moving dynamic content out, the hit rate reached 91–99.5%, plus savings of up to 47–62% on Opus 5.5 and Sonnet 5.5.

👤 For your own agents: check if the first request reads the cache; move the date, directory, and tools out of the cached prefix; read files in parallel; remove unnecessary sub-reviewers. At the same price per token, a model can be cheaper — it consumes fewer tokens. Cloudflare calculated its own quality metrics.

Source 1: https://blog.cloudflare.com/obsessive-inference-optimization/