GPT-6 Luna

GPT-6 Luna is a screening seat to run on every pull request: a measured review cost $0.0028, and it caught 67% [53%–80%] of the known bugs at rank 1 of 17.

The limit is recall. It measurably beats 4 ranked models and 0 measurably beat it, but its interval is wide. For a change where a miss is expensive, add a higher-recall seat rather than trusting Luna alone.

What makes it pleasant to run is the quiet. It averaged 0.7 [0.3–1.1] false alarms per clean PR, and 97% of its catches sat on the defect's own lines, so a comment usually points at the right spot. It also averaged fewer false alarms than Claude Sonnet 5.5, whose interval (2.0 [1.5–2.4]) does not overlap.

Claude Haiku 5.5 is its closest-priced rival from another maker: Luna lists at $0.10 / $0.50 per 1M tokens and Haiku at $0.10 / $0.50. Recall cannot be told apart, and a Haiku review cost 1.4× Luna's. Against GPT-6.1 Sol, a Sol review costs 14× Luna's; the benchmark cannot separate their recall, so whether Sol is worth that depends on the stakes, which the cost guide walks through.

A review takes 29s, fine for a CI check. The per-language and per-kind bars rest on a few fixtures each, too few to call a strength or weakness.