Spent an afternoon with GLM-5.3: Coding is indeed stronger, but don't switch yet

A friend spent the whole afternoon hyping up Zhipu's GLM-5.3, saying its coding ability jumped another notch. I've been using the Zhipu ecosystem for about a month now. When GLM-5.2 first went open source, I hooked up the API immediately. My impression then was that long context handling was solid, but it would occasionally get stubborn when writing code, making fixes a hassle. So when 5.3 dropped, my first reaction was to check how much fluff is in that '50% improvement' claim.