Does Claude Actually Read What You Paste? I Tested It to 250,000 Words

Оригинальный автор：Under_The_Prompt

Первоисточник：https://www.youtube.com/embed/2ON5xvzYp9Y

I built a 60,000-word document out of real Python documentation, hid one sentence in the middle that exists nowhere else, and asked Claude about it. Normal account. Screen recorded. Every answer exactly as it appeared including the ones where I was wrong. Short version of what happened: • Attach a file and it doesn't read it. It ran grep, printed the lines around the hit, and answered from those. Told explicitly to "read the entire document", it wrote a program to diff the file against Python's own docs instead. Correct both times — but it never read the text. • Paste the text and it does read it. It found the sentence at 20k, 60k, 120k and 240k words with the sentence dead centre, first try, and told me which section it was in. • It flagged a contradiction I planted 1,300 paragraphs away without being asked, and refused to invent a fact that wasn't there. • It caught a mistake inside my own hidden sentence — in every single run. I wrote that sentence and never noticed. • At 480,000 words the app accepted a 3 MB paste, showed the normal chip, and the content never reached the model. It said so, then answered correctly from its memory of the earlier chats. The rule I now use, by task: Need a fact? Attach. It'll search and give you the line. Need judgment? Paste it. That's the only way every word is in context. Bigger than a paste? Search first, paste the hits. Always ask "where?" and check the answer contains something only your document has. Otherwise you may be hearing memory, or search, or a guess. HONEST NOTES — please read these before quoting the results • One model, one day, one account, fourteen chats. These are observations, not a benchmark. • Model shown in the UI for every run: Claude Opus 5 High, on claude.ai, 18 September 2026, on a normal personal subscription. Cost: $0. • Not everything on screen was recorded live. The first set of runs (R1, R2, R5, R7, R8) happened before I started recording. What you see are re-runs (R1b, R5b, R7b, R8b) that produced the same results. Both sets of logs are linked above — compare them yourself. • T1 and T2 (the "attach and ask it to review" tests) were not screen-recorded at all. They appear as captions; the full text logs are in the repo. • The 120k, 240k and 480k documents repeat the same documentation two, four and eight times, with one needle. The needle stays unique, but the haystack is less varied than a real document of that length would be. • My account has cross-chat memory turned on. Answers in R2–R10 cited details that only existed in that run's pasted text, so those came from the document but a clean replication needs a fresh account, or memory switched off. • Token counts are estimates (characters ÷ 4). • The 3 MB paste limit is whatever the app does today. It may change.