Claude Haiku 5.5

Claude Haiku 5.5 is a cheap first seat: it catches 64% [50%–78%] of the known bugs for $0.0038 a review. It ranks 1 of 17, and a shared rank means the benchmark can't measure a gap between the models that hold it. It measurably beats 2 ranked models, and 0 measurably beat it. Its recall interval is wide, so this run can't place it precisely against the leaders in either direction.

Its closest match is GPT-6 Luna. Haiku lists at $0.10 / $0.50 per 1M tokens and Luna at $0.10 / $0.50, and the two cannot be told apart on recall. Luna averaged fewer false alarms (0.7 [0.3–1.1] against 1.4 [1.0–1.9] here, with overlapping intervals), and Haiku's measured review costs 1.4× Luna's. Haiku counted more input tokens and wrote more output, so matching list prices did not mean matching bills. On this data Luna is the tighter pick, and Haiku earns a seat as the second opinion from a different maker.

Expect some false alarms to dismiss, and some misses: "Fix DataClassEq comparator" and "Add weight option" are among them.

Do not use it alone on risky changes. Pair it with a high-recall seat, as the single model versus panel guide describes, or see how it trades against Claude Opus 5.5 on the compare page.