But it also frequently uses massively more tokens for the same task compared to other models which kind of negates the speed.

Wallclock time matters and GLM 5.3, even when it is considerably slower (~30 tokens right now vs 100+ on DeepSeek 4) it is quite frequently faster on the same set of tasks overall. Deepseek 4 seems to do the 'Oh, wait' thing just about forever and has a tendency to find irrelevant rabbit holes that it then spends a massive amount of tokens on.