Iris pits GPT against Gemini as prosecution and defense before Opus weighs the evidence—adversarial review at $4 per PR.
Iris Code Review, an AI reviewer that uses an adversarial multi-model architecture. When Opus finds a potential problem, it spawns two sub-agents—one to argue the concern is real, one to argue it isn't—then weighs the evidence before deciding whether to comment.
Opus 4 leads the review with access to sub-agents and general tools. When it spots a potential issue, it calls propose_possible_problems, which spawns GPT-5 as "prosecution" and Gemini 3.0 Flash as "defense." Each investigates and presents evidence. Opus evaluates impartially and decides whether to include the finding. Sub-agents read from an "AI reading room" filesystem containing code, PRs, Sentry issues, and tickets.