Dual-model architecture

Why Opus reasons and Sonnet executes.

Ghost-hunter uses two Claude models for two different jobs.

Claude Opus — the reasoner

Opus generates and ranks hypotheses, decides what command to run next, and writes the final root-cause report. Opus is slow and expensive — Ghost-hunter calls it sparingly, only when reasoning is required.

Claude Sonnet — the executor

Sonnet validates command syntax, compresses raw gcloud / aws output into structured findings, and answers cheap classification questions ("does this output confirm hypothesis 3?"). Sonnet is fast and cheap — most of the round-trips are Sonnet.

Why split

A single model would either be too expensive (everything on Opus) or too dumb (everything on Sonnet). The split lets Ghost-hunter spend Opus tokens only where reasoning quality matters and Sonnet tokens everywhere else.

Typical investigation cost: $0.15–$0.40 in Anthropic API spend, the bulk of it Opus.