Claude Opus 5.5

Pick Claude Opus 5.5 for the seat that guards risky changes: it caught 90% [81%–97%] of the known bugs and is measurably above 9 of the 17 ranked models on recall; 0 are measurably above it.

That lead is over the middle and bottom of the table, not over the rest of the top group. GPT-6.1 Sol and Opus can't be told apart on recall, and the head-to-head shows how differently they bill. A review on Opus cost $0.1525, 3.8× what a Sol review cost, and took 33s. Part of that gap is the list price ($4.00 / $20.00 per 1M tokens against Sol's $2.00 / $10.00), and Opus also wrote longer reviews.

It is not quiet. It raised more false alarms than Sol: 1.7 [1.1–2.3] per clean PR against 0.9 [0.4–1.5], and the intervals overlap, so treat the difference as a lean, not a ruling. When it does catch a defect, 89% of its catches land on the defect's own lines.

Two cautions:

• One of the three judges is from Anthropic, so its vote is dropped for this model. The other two still score it.
• Its misses include "perf(gitdiff): optimize inline diff highlighting using cache". Slices by bug kind or language are too small to say where it is strongest.

Use it as the high-recall seat on a panel and put a cheaper model first. See choosing a model for how to split seats.