Gemini, Gemini Notebook (formerly NotebookLM), ChatGPT and Claude can summarize a long video or a report of hundreds of pages, but each reads differently: some only a transcript, others only a PDF’s text and not its graphics. Free plans may not fit a 300-page document in working memory. Published limits help choose a tool and judge when a summary is unreliable. The root issue is the context window: how much text a model holds at once.
- Context by plan: Google lists 32,000 tokens for free Gemini, 128,000 for AI Plus, 1 million for AI Pro/Ultra; 300 pages would approach 200,000 tokens by its ratio.
- Video: Gemini Notebook imports captions, not images, so silent on-screen tutorials lose the point.
- PDFs: standard ChatGPT retrieval is text-only; Claude reads text and images only to 100 pages.
- Quiet failure: past the window, models can omit details while sounding complete.
Context windows and the 300-page report
Google translates 1 million tokens to about 1,500 pages; at the same ratio 300 pages would be near 200,000. That estimate varies with density but shows a 300-page report overflows both lower tiers. Google concedes that beyond the window answers may not account for all content or may lose connections and details, especially when a question needs many scattered data points. It warns that even a fitting but oversized file can omit relationships or details.
Google says responses may “no tener en cuenta todo el contenido” or lose details, and that a large file can produce a response “que omita relaciones o detalles del contenido.”
Gemini Notebook is better for long files, but only as text. It accepts public videos with creator or automatic captions, rejects videos without speech, and may fail on videos from the last 72 hours. There is no duration cap unless subtitles exceed 500,000 words. Each source allows 500,000 words or 200 MB; free users get 50 sources per notebook. Sources can be selected, and Google advises specific chat requests over a generic “resume esta fuente.” Caption errors carry into the summary. It also links a step-by-step NotebookLM podcast guide.
What Gemini, ChatGPT and Claude actually read
Whether AI can “see” a full YouTube video depends on the tool, and it almost never sees it like a person. Gemini’s main app can use public YouTube information with Save activity enabled. Google already allowed free YouTube summaries in Gemini in 2024. Uploading your own video caps at 2 GB per video and 5 minutes total without a subscription, or 1 hour with AI Pro or Ultra.
ChatGPT’s FAQ sets 512 MB per file and 2 million tokens per text document; free users get three uploads daily. Except in Enterprise, PDF retrieval is text-only and images are discarded. Anthropic’s July 2026 help allows 500 MB per file, 20 files per chat and PDFs to 1,000 pages. Claude reads text and images only in PDFs of 100 pages or less; from 101 to 1,000 it reads text only. Projects cap files at 30 MB, and paid plans activate retrieval near the context limit. If a 300-page report keeps conclusions in tables or scanned charts, neither tool guarantees reading them on general plans.
Where summaries fail, and how to check them
Long summaries fail by omission and invention; invention is harder to spot. FABLES (arXiv, 2024) used human readers on 3,158 claims from 26 book summaries. Most fidelity errors involved events and character states, often needing indirect reasoning to detect, and models weighted endings more heavily. Lost in the Middle (2023) found long-text answers were better when evidence sat at the start or end and worse in the middle. An AI summary tends to sound complete even when it is not; omission leaves no trace, and no one misses a chapter they did not know existed.
- Choose by format: transcripts for narrated video; for graphics, confirm your plan analyzes them.
- Split the material: request chapter or block summaries; Claude recommends this for big files.
- Request literal citations: Anthropic advises extracting verbatim quotes first for documents over 20,000 tokens.
- Ask what is missing: “¿Qué apartados no has usado?” or “¿qué dice el capítulo 7?” exposes gaps better than rereading.
The promise of “resume este PDF” is well sold and poorly explained. A summary is an orientation aid, not proof of full reading; the burden stays with the reader to pick the right tool, split the file, or ask what was skipped.