Compare two PDF versions: what a text comparison can tell you

Open Compare PDFs and add two ready PDFs. The first is called Document A and the second Document B, in the order you add them. So add the older version first.

The default model is Claude Sonnet 4.6, at 5 Gems per message. The answer has four parts:

• What each document is. A line on each, so you can see it read the right files.
• A table of the differences that matter. Columns: clause or topic, A, B.
• Added or removed. What appears in one document and not the other.
• Numbers or dates that changed. Fees, quantities, deadlines.

It ends with a one-line verdict on which differences need attention.

This is a real limit in the product. The free plan reads about 8,000 tokens per message, and each PDF counts as about 6,000. Two PDFs add up to far more than 8,000, so Compare is blocked, with a message that a paid plan reads more.

Paid plans read 16,000 tokens per message. On GPT-5 Pro or Claude Opus 4.7 the limit is 32,000, and a message costs 25 Gems. Each PDF must also be within about 5.7 MB.

See pricing for plans. On the free plan you can still run each version through Explain PDF and compare the two readings yourself.

Compare works on extracted text. That is a real strength for wording and numbers, and a blind spot for everything else. It can miss:

• Formatting. Bold, strikethrough, a moved heading or a changed layout.
• Moved signatures. A signature block that sits on a different page.
• Images and stamps. A logo, a seal or a stamp is not text.
• Handwritten edits. Pen changes on a printed page are not in the text.
• Anything in a scan. A scanned page has no text layer, and there is no OCR.

It is also not a redline, which would mark every changed word on the page. And it is not legal advice. Check every reported change against the page in both files before you rely on it.

The examples in this guide are fictional. They are not real files and not real model output.

Imagine a service contract. Version 2 raises the monthly fee, shortens the notice period from 60 days to 30 days, and adds auto-renewal. A good Compare answer puts those in the table:

• Monthly fee: A states one amount, B states a higher one.
• Notice period: A says 60 days, B says 30 days.
• Renewal: A has no renewal clause, B renews automatically.

The verdict line would flag all three. The notice change and the renewal clause work together: with 30 days to cancel and a contract that rolls over by itself, a missed date costs another term. That is the kind of thing worth a second look, and it is why you open both files at the named clauses.

For reading one contract in plain words, see how to understand a contract with AI.

Yes, and quotes are a good fit, because the question is usually about terms, not wording. Another fictional example: two quotes for the same office furniture order.

Quote A promises delivery in 3 weeks and includes assembly. Quote B is cheaper on the line items, promises delivery in 6 weeks, and says nothing about assembly. A comparison table puts delivery time and assembly side by side, so the cheaper total stops looking like the better deal.

Ask the model to compare like for like: what each quote includes, delivery terms, payment terms, and what one has that the other lacks. Quotes from different suppliers use different layouts, so check that each price it lists is the total you mean, not a unit price.

Treat the table as a to-do list, not a verdict. Work through it in this order:

• Open both files at the named clause. Put Document A and Document B side by side and find the clause the table names.
• Match the wording. Ask for the old and new wording to be quoted, then find each quote on the page. A quote you cannot find is a warning sign.
• Recheck the numbers. Read each fee, quantity and date in both files yourself. A model can misread a figure in a dense table.
• Look for what the table did not list. Ask whether anything else differs, then skim the headings of both files for sections that exist in only one.
• Check the parts text cannot show. Signature pages, stamps and handwritten notes need your own eyes.

A fictional example: the table says the notice period moved from 60 days to 30 days. You open both files at the termination clause, find the two sentences, and see that the shorter period applies to only one party. The table said nothing about that, and now you have a question for the other side.

Most bad comparisons come from the setup, not the model.

• Adding the files in the wrong order. If the newer file goes first, every added clause looks removed. Add the older version first.
• Mixing a digital file with a scan. The scan has no text layer, so the model has nothing to compare it with. Use digital copies of both.
• Comparing whole files when one section matters. Long files may not fit in what the model reads. Cut both to the relevant sections first.
• Trusting a short answer. "No major changes" is not the same as no changes. Ask for a complete list, including small reworded sentences.
• Mixing up units in quotes. One supplier may price per item and another per order. Ask the model to say which each price is before you read the totals.

If the answer surprises you, ask the same question again with a narrower scope, such as one clause. Two matching answers are worth more than one long one.

Two long contracts may not fit in what the model reads, and each file must stay within about 5.7 MB. Cut both down to the relevant sections first.

Use the free Split PDF tool. It runs in your browser, so nothing is uploaded and no Gems are spent. Pull out the same sections from each version, such as fees, term and termination, then compare those.

Both files go to the model you picked, and they are sent again with every follow-up, so keeping them short also keeps less of the contract leaving your device.