A 50-PR benchmark says GPT-5.6 Luna reviews code at 28x less cost than Astra, at 74% precision against 96%

Entelligence published a benchmark on September 14 that answers a question a lot of teams are guessing at right now (Entelligence). They ran GPT-5.6 Luna and GPT-6 Astra over 50 real public PRs, ten each from Cal.com, Sentry, Discourse, Keycloak and Grafana. Identical prompts. A dual-judge scheme where GPT-6 Astra and GPT-5.6 Sol both had to agree before a finding counted as a real bug.

Luna: $0.0041 per review, 69 verified bugs, 74% precision. Astra: $0.113 per review, 92 verified bugs, 96% precision.

That's a 28x cost difference for about a third more findings and 22 points of precision. The lazy read is "cheap model is worse, pay for the good one." The overlap data refuses that read. Luna caught 25 bugs Astra missed. Astra caught 48 Luna missed. These are not nested sets. They're two partially overlapping circles, which means running both finds more than running either.

The security split is where this turns into a routing rule. Of 24 security bugs in the corpus, Astra found 19 and Luna found 9. The authors won't let Luna review auth or permission code alone, and the number is why.

So the policy writes itself: cheap model on the general pass, frontier model on anything touching authentication, authorization, session handling or secrets. That's a path-based router in your CI config, not an architecture project. Twenty lines of YAML matching auth/, permissions/, middleware/ and routing those diffs to the expensive model while everything else goes to the cheap one.

I've been running a single model over every diff and eating the bill. Looking at those numbers, I was paying frontier prices to have a frontier model tell me about a missing null check.

Put this next to Ericsson's industrial result from the same week (arXiv 2609.15877). They built a multi-agent code reviewer, ran it under the Design Science Research Process, generated 200+ issues across several commits, and had their own developers manually validate every one. 96% accuracy in correctly identifying issues, with 69% of correct findings rated important and about 33% severe enough to be must-fix. That last number is the one I'd quote to a skeptical team lead. A third of what the reviewer found had to be fixed. Not nitpicks about naming. Real defects.

Published industrial precision numbers on agentic review are rare. Most of what circulates is vendor marketing or a blog post about someone's weekend. This is a named company, validated by the company's own engineers, with a methodology section.

The cost story connects to the Koa story below in a way that took me a minute to see: capability is detaching from price, and it's detaching from pretraining budget at the same time. Two different axes, same direction.