Why are some GPT 5 API requests costing three times more than others?

We've been using the GPT 5 API for document processing and noticed that some requests are costing two or three times more than others even when the documents are roughly the same length. At first I assumed it was just the input size but after looking at the usage numbers, that doesn't seem to explain everything cause some requests use significantly more output tokens and I'm wondering how much of that is coming from reasoning versus the actual response. It's getting harder to estimate what processing a batch of documents will cost especially when the requests aren't all behaving the same way.