Sonnet 5.5 handles collection and coding, but falls short on review work

The Zenn article is a practitioner's report from hands-on comparison runs, not a vendor benchmark. The author built six known-answer test cases from past review comments and incidents, ran them headless with claude -p, and had two separate models grade the outputs after anonymization.

The article's practical point is that effort and model choice should be decoupled by task risk. The author maps this to a three-dial model: model, effort, and a separate prompt-level instruction to actually verify claims against sources. Tasks where an error would directly affect a user's decision or deliverable call for Opus at high effort or above, while collection, aggregation, exploration, and post-plan implementation run on Sonnet at medium. The default effort values differ by model - medium for Opus 5.5 and Sonnet 5.5, high for many others - so leaving settings untouched can silently run heavy work shallow.

The author also flags a limit: some review angles, including whether tests can catch regressions, whether reused old branches roll back later fixes, and whether parallel request limits are respected, were missed by every model and every verification instruction tested. Manually adding those checkpoints to a procedure or agent definition may be more effective than switching models.