Opus 4.6. My standard battery of questions included solving an ascii maze (20x20 grid) without using a script, using only "thinking" as a tool. It was the first model to be able to solve it. It was the first model that really appeared to be able to reason spatially.
Opus 4.6 for me as well. I had a serious bug in some legacy software I've been stuck with maintaining, together with a few other people who originally wrote the software. We've all been trying to solve this bug for literally 10 years or more. None of us have been able to. I've personally spent hundreds of hours on it, thrown it at every previous LLM. Opus 4.5 came up with a workaround that prevented our software crashing, but didn't solve it. Opus 4.6 was the one that actually solved it. It did it by modelling a state machine of the software that was calling our software and triggering the bug, and it found the one state where we weren't correctly sending data back.
