Token Benchmarks
Why tokens matter, and what TrimDown saves you.
The saving is the second reason to use TrimDown. The first is the Word document you get back — see Home.
Save Money
A PDF page sent to a model is billed as an image and as extracted text together, 1,500 to 3,000 tokens per page, per Anthropic's own PDF support documentation. Output tokens cost roughly five times as much as input tokens across Claude, GPT and Gemini alike. TrimDown's Markdown cuts both sides of that bill. Below are five synthetic benchmark documents, the real savings they produced, and the files to check our numbers yourself.
Save Time
There's a second saving beyond the bill: time. Fewer tokens per page means a document that would blow through a model's context window as a PDF often fits whole as Markdown — no splitting a report into chunks, no "summarize part 2 of 3." and no session on hold. And because TrimDown shows you the Markdown before you send it anywhere, you can proofread what the model is actually going to read, not what you assume it's reading. That step catches mistakes before they leave the building: PDF interpretation errors are common and delivered with total confidence. One writer gave ChatGPT, Claude and Gemini the same 47-page report and caught all three making things up. Another documented an AI that misread a bar chart, flipped a label, and reported a 14% decline as an increase — stated as fact, no hedge. Catch that kind of error in the Markdown before it's in an email to your boss or your customer, not after.
In a common session, every file is read few times.
The savings multipy!
Our Sample Files Savings
(1) All tests were run on Claude Sonnet 5, Opus 5 and Fable 5.1. Add 2 tokens for each number for Fable 5.1.
All numbers are actual numbers of the input tokens reported by Claude through Claude API.
What actually goes to the model
A Word file or a PDF is mostly pictures. When TrimDown converts one, the pictures are written into a folder beside the Markdown so the preview can draw them and so an export can put them back. That folder stays on your Mac. Only the Markdown goes to the model.
The folder can look larger than the file it came from. That is not TrimDown adding anything — a Word file is a compressed archive, and the same pictures take more room unpacked than packed.
Size on disk is not what you pay for. You pay for tokens. One chart costs about 646 tokens sent as a page image and about 49 tokens sent as a table of numbers — and the table is worth more, because the model can read the values.
Run the tests yourself
Test 1, input tokens. The quick way: open a new conversation with your model of choice, attach the original PDF and read the input-token count your interface shows, then start a second fresh conversation and paste in the TrimDown Markdown version and compare. Use a fresh conversation each time. The exact way: run measure_tokens.py from the downloadable set, which prints page count, image tokens, text tokens, PDF total, Markdown tokens and the saving. For an exact Claude count, send both payloads to Anthropic's Messages Count Tokens API.
Test 2, output tokens. Ask your model for a formatted report with charts the expensive way, a chart description it must render in prose. Then ask for the same report in Markdown with each chart as a Mermaid diagram, and export it to Word or PDF with TrimDown, so the chart renders locally instead of being described by the model. Compare the output-token counts.
Methodology
Validated Efficiency
Our methodology focuses on aggressive reduction of non-semantic overhead. By converting complex binary PDF structures into clean, logically structured Markdown, we strip away redundant visual formatting while preserving the core informational density required for LLM reasoning.
What the documentation says
"A PDF is billed as page images and extracted text together, not one or the other."
"ChatGPT does not read your PDF in full. It indexes and retrieves parts."
"Prompt caching changes the cost arithmetic in your favor on all three leading engines."
Our benchmark notes; not modeled in the figures above
Honest limits — please read before quoting our numbers
Is this Claude's own tokenizer?
Yes. Every figure on this page is an exact input-token count returned by Anthropic's Messages Count Tokens API, not an estimate from a third-party tokenizer.
Do real documents save as much as these synthetic ones?
Yes, but the numbers will vary, depending on the document. Our own PRD (Product Requirement Document) with 22 screenshots saved 85.1% with Sonnet 5, Opus 5 and Fable 5.1, and 88.5% with Haiku 4.5. Two other files, a 27-page analyst report and an 8-page Hebrew lease contract, saved 63.3% and 60.5%, both below every synthetic figure above.
Is the round-trip export figure published?
Not yet, on purpose. TrimDown draws Mermaid diagrams in the live preview and exports them into the Word document as real pictures — that part works today. What we have not yet measured is the token cost of the whole round trip: document in, Markdown out, answer back, document out. We will publish that figure when we have measured it, and not before.
Does language change the saving?
Yes, but less than page count. Hebrew and other non-Latin scripts tokenize at roughly 1.8 to 2.6 characters per token against 3.5 to 4.8 for English, which is why our synthetic Hebrew test file shows 78.8% while the denser real Hebrew contract shows 60.5%.
Is Markdown always cheaper?
No. For dense tabular data, other formats can beat it. We are claiming an advantage for documents, not for every payload.