Let a cheap model fix real bugs overnight, and most of its PRs didn't survive review

This is my own setup, my own repos, and I'm a bit stuck on what to do with the result. An audit turned up some serious bugs. Instead of fixing them myself I let a cheap, fast model (DeepSeek, during an API discount window) grind through them overnight. Round 1 was 14 PRs. I had a stronger model (GPT-6.1 Sol) re-review every one of them. 6 got merged after all CI passed. I closed the other 8. The reasons were things like a fix that broke other valid inputs, a test that needed a tool our CI doesn't have, a bypass that still worked, a key test that never even ran in CI, and a new test that couldn't catch the bug even when I deliberately broke the code. So for round 2 I flipped it. A stronger model designs each fix, another one reviews the design, and the cheap model only implements. That's 7 PRs so far and none approved yet. Where do cheap-model fixes fail for you? And what split between planning, implementing and reviewing has actually worked?