CodeRabbit’s reviews on Sonnet 5.5 cost about 60% less

On 13 hard known-bug cases from real open-source pull requests, Claude Sonnet 5.5 caught 6 and Sonnet 5 caught 4, at about the same precision. CodeRabbit ran the test through its own review pipeline and is already moving simple and moderate reviews to the new model.

The Claude calls cost $0.47 per review, against $1.16 for Sonnet 5. Anthropic’s own launch promise is only “up to 30% less per task.” On 44 pull requests, the average review took 6:33 instead of 13:31.

Both models have the same price list, so the whole gap comes from tokens. On each review call, Sonnet 5 reads more than twice as much and writes about four times as much.

CodeRabbit admits 13 cases is a small set, and on the hardest ones Opus 5.5 is still the stronger reviewer. Turns out the cheapest upgrade was a model that argues with itself less.