claude vs gemini for a genuinely big document, who holds up past like 200k tokens
ok so undergrad here doing a lit review, i've got a pile of pdfs and i keep hearing gemini has the biggest context but claude 'stays coherent'. has anyone actually stress tested this on real long docs not a needle-in-haystack toy. like does gemini's huge window actually reason over the whole thing or does it just retrieve a line and lose the plot. burned an afternoon on this already and got mixed results, want to know if im doing it wrong
ran claude vs gemini 3.1 pro on a ~340k token contract set last month. gemini technically holds the whole thing but recall past ~200k got mushy, it'd cite a clause from the wrong doc. claude with a retrieval pass over chunks beat raw-context gemini on accuracy even though gemini 'saw more'. big context window isnt the same as using it. what are you actually loading
ok so this matches what i saw exactly, recall goes mushy past 200k and it grabs from the wrong doc. did you rerank or just stuff it all in? the reranker is what saved me, raw big-context was worse than a good 30k window