How many tokens is a 10-page PDF?

A typical text-based page holds 400–600 words, so a 10-page PDF is roughly 4,000–6,000 words, or about 5,500–8,000 tokens for GPT models (English averages ~1.33 tokens per word). Dense legal or academic pages run higher. Upload your actual PDF above for the exact count, page layout and formatting change the total.

How much does it cost to summarise a PDF with ChatGPT via the API?

Only the input tokens you send are billed for reading the document; the summary you get back is billed as output. A 24,000-token PDF costs about $0.03 of input on GPT-5 (Rs 9 at Rs 300/USD) or $0.06 on GPT-4o. Add the reply: 800 output tokens on GPT-5 is another $0.008. Enter your expected reply length above to see the total.

Will my document fit in Claude's context window?

Claude Sonnet 4.6 and Opus 4.8 accept up to 1,000,000 tokens, so almost any single document fits in one call. The tool marks each model ✓ if the whole document fits, or shows how many chunks you'd need otherwise, each chunk reserving 4,096 tokens for the reply. GPT-4o (128k) and DeepSeek (64k) fill up much sooner than the million-token models.

How do I count tokens in a Word document without an API key?

Drop the .docx here. The tool unzips it in your browser (a .docx is a ZIP of XML), reads the document body, and counts tokens locally with OpenAI's tokeniser. No API key, no upload, no signup. The same works for PDF, plain text, Markdown, CSV, JSON and source-code files.

Why does the same file cost different amounts on GPT-5 vs Claude?

Two reasons. First, per-token prices differ, GPT-5 input is $1.25 per million tokens while Claude Sonnet is $3.00. Second, each model tokenises text with its own vocabulary, so the same document is a slightly different number of tokens on each. The cheapest model here is usually Gemini Flash or Llama; the priciest is Claude Opus.

Are the token counts exact?

For OpenAI models (GPT-5, GPT-4o) they are exact, the tool runs tiktoken's o200k_base tokeniser in your browser, the same one OpenAI bills with. Claude, Gemini, Llama, DeepSeek and Mistral counts are estimates using each vendor's published character-per-token ratio, because their exact tokenisers aren't available client-side without an API call. Estimate rows are labelled 'est'.

Do my files get uploaded anywhere?

No. Every step, PDF text extraction (pdf.js), DOCX unzipping (JSZip), and tokenisation (gpt-tokenizer), runs inside your browser tab. The document bytes never touch our server or any third party. Only the pdf.js worker script is fetched from a CDN; your file stays on your device.

What about scanned PDFs with no text layer?

The tool reads a PDF's embedded text layer. A scanned or photographed PDF has images, not text, so it returns zero tokens and flags the file. Run it through an OCR tool first to get selectable text, then paste that text here to count it.

Where do the prices and the rupee conversion come from?

Per-model USD prices and context-window sizes are reused from the maintained AI API Cost Calculator dataset on this site, which cites each vendor's pricing page inline. The USD→LKR figure defaults to the Central Bank of Sri Lanka indicative rate and is editable, so you can drop in today's rate. All sources were last verified on 2026-07-11.

AI · Developer tools

AI Document Token & Cost Calculator

Drop a PDF, Word or text file and see its exact token count, whether it fits in each model's context window, and what it costs to send to GPT-5, Claude, Gemini, Llama and more, in USD and rupees. Parsed entirely in your browser: no upload, no API key, no signup.

By Induwara Ashinsana· Co-Founder, Ryzera TechnologiesUpdated Jul 11, 2026

How it works

This tool answers one question the paste-text token counters can't: "I have a file, what will it cost to run through each model, and will it fit?" It combines client-side file parsing, real tokenisation, maintained pricing, and context-window fit in one place. Nothing you upload leaves your device.

Extract the text, in your browser.PDFs are read with Mozilla's pdf.js (the page text layers are concatenated); Word .docx files are unzipped with JSZip and the document body is stripped from its XML; plain text, Markdown, CSV, JSON and source code are read directly as UTF-8. Scanned PDFs with no text layer are flagged, not guessed.
Count tokens per model family.For OpenAI models the count is exact: the tool runs tiktoken's o200k_base tokeniser (via gpt-tokenizer), the same byte-pair encoding OpenAI bills with. Claude, Gemini, Llama, DeepSeek and Mistral use each vendor's published character-per-token ratio: ~3.5 chars/token for Claude, ~4 for Gemini/Llama/Mistral, and ~3.33 for DeepSeek. Those rows are labelled an estimate.
Check context-window fit.A document fits in one call when its token count is at or below the model's context window. If it doesn't, the number of chunks is ceil(tokens / (contextWindow − 4096)), reserving 4,096tokens of each window for the model's reply.
Price the input. Input cost is tokens ÷ 1,000,000 × inputPricePerMillion. Adding an expected reply length adds outputTokens ÷ 1,000,000 × outputPricePerMillion for the total. Every USD figure is converted to LKR with your rate.

Prices, context-window sizes and the default exchange rate are reused from this site's maintained AI API Cost Calculator dataset, where every per-million figure carries an inline source URL. The arithmetic is cross-checked by an independent per-thousand-token formula and a set of hand-computed worked examples that must reconcile before release, the same verification approach as the tax calculator.

Worked examples

A 42-page contract PDF (~18,000 words ≈ 24,000 tokens)

Extract text from the PDF → 24,000 GPT tokens (exact, o200k_base).
GPT-5 input: 24,000 ÷ 1,000,000 × $1.25 = $0.0300 = Rs 9.00
GPT-4o input: 24,000 ÷ 1,000,000 × $2.50 = $0.0600 = Rs 18.00
Claude Sonnet input: 24,000 ÷ 1,000,000 × $3.00 = $0.0720 = Rs 21.60
Gemini 2.5 Flash input: 24,000 ÷ 1,000,000 × $0.30 = $0.0072 = Rs 2.16
Fit: 24,000 ≤ every window (all ≥ 128k) → single call on all. Cheapest: Gemini Flash.

A 250,000-token research corpus (many files)

Combined text tokenises to 250,000 tokens.
GPT-4o input: 250,000 ÷ 1,000,000 × $2.50 = $0.6250 = Rs 187.50
GPT-5 input: 250,000 ÷ 1,000,000 × $1.25 = $0.3125 = Rs 93.75
Fit vs GPT-4o's 128k window: ceil(250,000 ÷ (128,000 − 4,096)) = ceil(2.02) = 3 chunks.
GPT-5 (400k) and Claude Sonnet (1M) take it in a single call.

Edge case, a scanned PDF, then a boundary check

A scanned PDF has no text layer → 0 tokens extracted, file flagged for OCR (no silent guess).
Boundary: a document of exactly 128,000 tokens fits GPT-4o's 128k window → 1 call.
One token more (128,001) no longer fits: ceil(128,001 ÷ 123,904) = 2 chunks.
Zero-length input costs $0.00 on every model, no NaN, no divide-by-zero.

Frequently asked questions

Sources & references

Tokenisation ratios and the model roster were last verified on 2026-07-11. Per-model prices are maintained in the shared AI API Cost Calculator dataset and reviewed quarterly.

Comments & feedback

Spotted a bug or want an improvement? Tell us, our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want another file type supported?

Email me at [email protected] , most fixes ship within 24 hours.

AI Document Token & Cost Calculator

Models to compare

How it works

Worked examples

Frequently asked questions

Sources & references

Related tools

Tool-Use Token Cost

Tokens to Words

AI Cost Per Word

Comments & feedback