The Caps Are Measured in Different Units
A comparison that says "ChatPDF: free tier available; Humata: free tier available" tells you nothing, because those two sentences describe completely different products. One resets every morning. The other runs out permanently after roughly four papers.
There are four units in play, and the right tool depends entirely on which one matches how you read:
- Documents per day — resets daily. Fine for steady reading, useless for a weekend literature sprint.
- Pages for the lifetime of the account — a one-time allowance. Once spent, you pay per page.
- Sources per project — generous overall, but forces you to split a big review across notebooks.
- Unlimited on the core action, limited on the fancy one — summaries free, autonomous agents capped.
Sort the tools by that and the choice makes itself.
What Each Free Tier Actually Gives You
Elicit Most generous free tier
The surprise of this comparison. Elicit's own pricing page lists the free plan as including unlimited search across more than 138 million papers, unlimited summaries across as many papers as you want, and unlimited chat with papers with full-text access, plus source display and Zotero import. What is capped is "limited usage for Research Agent and Research Reports" — the autonomous multi-step features, not the summarising.
For the specific job of "I have 40 papers and I need to know which five matter," that is the best free offer on this page by a wide margin. Paid Pro is $49/month (billed annually at $588), which tells you the free tier is not a rounding error they forgot about — it is deliberate positioning.
NotebookLM 50 sources per notebook
Google's free tier is specific and, per Google's own support documentation, unusually roomy: 100 notebooks per user, 50 sources per notebook, 50 chats per day, and 3 audio overviews per day. Plus raises that to 100 sources and 200 chats; Pro goes to 300 sources and 500 chats.
Fifty sources in one notebook is a real literature review. The constraint that bites is not the count but the shape: everything you want compared has to live in the same notebook, so a 120-paper review means splitting into three and losing cross-notebook questions. The daily 50-chat cap is the other thing to watch — that is 50 questions across all your notebooks, and interrogating a paper properly burns through them faster than you expect.
ChatPDF 2 documents per day
ChatPDF's free plan lets you analyse 2 documents every day. The file ceiling is generous — up to 2,000 pages or 32 MB per file — which is worth stating plainly because most articles assume the page limit is the problem. It is not. A 2,000-page allowance covers any journal article ever written, plus most theses. The bottleneck is the number two, and it resets tomorrow.
That makes ChatPDF good for a steady habit (two papers a day, every day, is a strong reading pace) and bad for the night before a supervision meeting.
Humata 60 pages, then it stops
Read this one carefully, because it is the free tier most likely to strand you mid-task. Humata's pricing page describes the free plan as "basic features for up to 60 free pages of use" — a lifetime allowance, not a monthly refill. A typical journal article runs 10 to 15 pages, so the free plan is roughly four or five papers in total, ever. After that it is $0.02 per page, or $9.99/month for the Expert plan with 500 pages included.
Nothing dishonest about it — the number is published clearly. But "free tier" and "free trial" are being used interchangeably here, and only one of those descriptions is accurate.
Semantic Scholar TLDR Free, no catch
The one with no tier at all. Semantic Scholar is a free academic search engine, and its TLDR feature generates roughly 20-word summaries of a paper's objective and results, available in beta across nearly 60 million papers in computer science, biology and medicine.
Twenty words is not a summary you can work from — it is a triage signal, the thing that tells you whether to open the PDF at all. Used as the first pass before any of the tools above, it is the single most efficient step in this entire article, and it costs nothing and caps nothing. Its blind spot is coverage: outside CS, biology and medicine, expect gaps.
General assistants: Claude, ChatGPT, Gemini
All three accept PDF uploads on their free plans, and for a single paper they are strong — you can ask follow-up questions in a way that a fixed summary cannot answer. The trade-offs are the ones covered in our comparison of the three assistants: daily upload caps that vendors do not publish as stable numbers, and no built-in link to the literature around the paper. If the paper is unpublished, confidential, or under review, keep it off all three and use a local model on your own machine instead.
A Free Workflow for a Reading List of 30 Papers
Combining free tiers beats maximising any single one. A sequence that stays inside every cap:
- Triage with Semantic Scholar TLDR. Twenty words per paper, unlimited, no account friction. Thirty papers become the eight worth real attention.
- Screen the eight in Elicit. Unlimited summaries and full-text chat on the free plan. This is where you decide which four you actually need.
- Load the four into one NotebookLM notebook. Well inside the 50-source limit, and now you can ask questions that cross papers — where they disagree, which methods overlap.
- Read those four properly. No tool replaces this step.
If you paste text out of a PDF rather than uploading it, two problems arrive with it: hard line breaks at the end of every visual line, and invisible formatting characters. Our guide on removing line breaks from PDF copy-paste covers the first, and AI Text Cleaner handles the second. Before pasting a long paper into any chat window, Token Counter tells you how big it actually is — a 12,000-word paper is a substantial chunk of a small context window, and silent truncation is why summaries sometimes ignore a paper's back half entirely.
What AI Summaries Get Wrong About Papers
Summarisers are reliable on the parts of a paper that are already written to be summarised — abstract, stated conclusions, headline numbers. They are weakest exactly where academic reading matters most.
Methods and limitations get compressed away. A summary will faithfully report that an intervention "significantly improved outcomes" and quietly drop that n=24, that it was unblinded, or that the effect vanished at follow-up. The validity of a paper lives in the parts a summary treats as detail.
Hedging gets flattened. "These results may suggest a possible association" reliably becomes "the study shows." That single transformation is responsible for a lot of confidently wrong literature reviews.
Never cite from a summary. If a claim is going in your work, open the paper, find the sentence, and check the surrounding paragraph. A summariser is a tool for deciding what to read, not a substitute for having read it — and if your institution has an academic integrity policy, that distinction is the one it turns on.
Comparison Table
| Tool | Free Cap (and its unit) | Best For | Limitation |
|---|---|---|---|
| Elicit | Unlimited summaries & full-text chat; agents limited | Screening a large reading list | Pro jumps to $49/mo for the agent features |
| NotebookLM | 50 sources/notebook · 100 notebooks · 50 chats/day | Comparing papers against each other | No questions across notebooks; 50 chats/day is shared |
| ChatPDF | 2 documents/day · up to 2,000 pages or 32 MB per file | A steady two-papers-a-day habit | Useless for a deadline sprint |
| Humata | 60 pages total — lifetime, not monthly | Trying it on one or two papers | ≈4 papers and it stops; then $0.02/page |
| Semantic Scholar TLDR | No cap — free platform | First-pass triage of 30+ papers | ~20 words; beta, ~60M papers in CS/bio/med only |
| Claude / ChatGPT / Gemini | Daily upload limits, not publicly fixed | Interrogating one paper in depth | No literature context; unsuitable for confidential drafts |
Which One Should You Use?
- You have a long reading list and no budget: Elicit. Unlimited summaries and full-text chat on the free plan is not matched by anything else here.
- You need papers compared, not just summarised: NotebookLM. Fifty sources in one notebook, and the cross-document questions are the point.
- You want a fast triage pass before committing: Semantic Scholar TLDR, every time. Free, uncapped, and it saves you from opening 20 irrelevant PDFs.
- You read about two papers a day: ChatPDF fits that rhythm exactly and the file size ceiling will never trouble you.
- You are evaluating Humata: treat the 60 pages as a trial, not a plan, and decide before you spend them.
- The paper is unpublished or under review: none of the cloud tools. Run a local model, or read it yourself.