For LLMs (apart from Apple's foundation models) the M1 Max still holds its own for the most part.

The base M6 is faster (much much faster) than the M1 Max at prefill (compute bound).

The M1 Max outperforms the base M6 at token generation; this is memory-bandwidth-bound, and while the M1 Max can't actually saturate its memory bandwidth, it gets a lot closer with the GPU.

So I suspect the M6 is better if your AI tasks are decision-based or video/image based, and the M1 Max is better if you're running a coding LLM. Though prefill is a bit of an issue.