Faster inference would solve nearly all of the issues I have with AI models. I really liked working with Opus 4.8 and I would greatly prefer a 100x faster version of it, with 100x the tokens, at the same price, to any of the openai or anthropic models that have come out since.