Gemini 4 Carbon is the only model I used that fixed a super subtle memory corruption bug in a JIT compiler I'm working on, hiding as an instructions flag bug.
Astra 6.1 and Opus 5.5 spent days on it without fixing it. Gemini 4 Carbon fixed it in about 3 hours. Only Gemini Carbon decided after getting the same result over and over that maybe the bug it assumed it was, was actually something different and tried a different idea on its own initiative.
Maybe Fable might have got it, but probably not before I ran out of credits.
Gemini 4 is _really_ good at fixing bugs. Whether the OpenAI 6.2 models are that good remains to be seen.