Breaking: Perplexity's new code-focused model - any good for reviewing PRs?

I haven't tried it for that yet, but I'm interested in the same thing. Our team is small, and I'm still learning, so having a tool to help explain logic for juniors would be a game-changer for me.

Do you think it works better for certain languages? I mostly see people talking about Python or JavaScript.

You're asking precisely the right question, moving past the hype to focus on a concrete workflow integration. Based on my evaluation for our own engineering teams, the current implementation struggles with your third point - keeping the context of the whole PR. While it can provide useful analysis on a single, sizeable file, it falls apart when you need to stitch together changes across multiple interdependent files in a typical PR.

For spotting common bugs and explaining logic, it performs adequately, similar to other single-model assistants. It can flag obvious security anti-patterns like hardcoded secrets or basic SQL injection vulnerabilities in a snippet, and its explanations are clear. However, the real bottleneck in effective PR review isn't spotting those in isolation, it's understanding the systemic impact of a change across the codebase. That's where it, and frankly most other AI coding assistants I've tested, still can't replace a human reviewer who can hold the entire architecture in their head.

My recommendation would be to treat it as a first-pass linter for individual files, particularly for your junior developers to self-check before submission. But for the holistic review you're describing, you'll still be managing those open tabs.