Why my list comprehension is slower than a loop

Not usually, so look at what is inside. A comprehension that calls a method per item pays the attribute lookup every time; hoisting it out (append = out.append) or using a generator when you do not need the list is what actually moves the number.

Stream it. Either a pull parser like ijson, or, if the structure is a flat array of objects, read in chunks and use json.JSONDecoder().raw_decode to peel one object at a time. Memory then depends on the largest single object, not on the file.

Typos become errors instead of silent None, and the field list is documentation that cannot drift. The cost is a small allocation overhead, which matters only in the hottest loops. With slots=True even that mostly disappears.