When Cloudflare first blogged about its AI code reviewer in April, an average review cost a palatable $1.19. By the end of August, that figure had climbed to roughly $2.46 — without any significant architectural change. The coordinator was still dispatching work to the same specialist reviewers, and the same risk tiers still determined review intensity. Around mid-September, the team tasked its agents with finding where the tokens were going and cutting them aggressively. The result: with code quality set to high effort, an average review now costs $0.77, down 68% from the $2.43 averaged over the two most expensive weeks before the work began.
- About 40% of the savings came from changes and optimizations to the reviewer's own code; the remainder came from experimenting with which models are used and how much effort they apply.
- The optimization also improved reviewer performance: the median review now takes 1m 32s, down from 4m 09s previously.
- Cost tracking holds the mix of review tiers fixed, so a day with more large reviews doesn't skew the comparison — giving a like-for-like view of per-review spend.
- The team used its own agents to analyze token consumption and identify where the spend was concentrated before cutting it.
The takeaway for developers building agentic or multi-model pipelines is that inference cost is rarely fixed by architecture alone. A stable coordinator-and-specialists design still doubled in cost as usage grew, and the savings came from disciplined, line-by-line tuning of reviewer code plus deliberate model and effort selection. That combination delivered both a 68% cost reduction and a roughly 2.7x speedup on median review time — a reminder that profiling token usage and right-sizing model effort can meaningfully change the economics of AI-assisted code review without compromising quality.
Source: https://blog.cloudflare.com/obsessive-inference-optimization/