#tHow Many Tokens?

← Back to counter

How many tokens per page?

Updated · By · Methodology

The short answer

A standard, single-spaced text page is ~250 to 500 tokens. Most documents land near 350 tokens per page.

The number swings wildly based on what's actually on the page:

Page typeTokens per page (approx)
Double-spaced manuscript (12pt)150 – 200
Standard novel (11pt)300 – 400
Single-spaced business doc (11pt)400 – 500
Dense academic paper (10pt)500 – 700
Two-column journal article700 – 1,000
Code listing200 – 400 (less efficient per char)
Table-heavy report300 – 800 (varies wildly)

These are for OpenAI's tokenizer (cl100k_base / o200k_base). Claude counts run roughly 30-35% higher since the Opus 4.8 tokenizer update.

Why the range is so wide

A "page" is a layout unit, not a content unit. The same word count can be displayed as half a manuscript page or a third of a dense academic page. Three factors drive the spread:

Density. Words per page varies from ~200 (double-spaced, large margins) to ~700+ (single-spaced, narrow margins, small font). Token count scales linearly with words.

Content type. Tables, code, and formulas don't tokenize as efficiently as prose. A 500-character line of natural English is ~85-100 tokens. The same 500 characters of code with lots of symbols, indentation, and identifiers can be 140-180 tokens.

Whitespace. Most tokenizers treat each significant whitespace run as a token boundary. Heavily formatted text (markdown tables, bullet lists with deep nesting) generates more tokens than the same content as flowing paragraphs.

Worked examples

For a standard letter-size page at 11pt Times New Roman, single-spaced, with 1" margins:

For double-spaced manuscript pages (~250 words):

For a typical academic paper page (single-spaced, 10pt, with citations):

How to estimate from page count

The fastest mental math for a typical document:

OpenAI / GPT models: pages × 350 for normal documents, pages × 550 for dense academic content.

Claude models: Multiply the GPT estimate by 1.33.

So a 50-page report ≈ 17,500 GPT tokens / 23,000 Claude tokens. A 200-page textbook chapter ≈ 110,000 GPT tokens, fits in Claude's 200k window, fits in GPT-5's 400k window with room.

When the math breaks

These heuristics fall apart for specific content types:

Scanned PDFs. A scan is *images*, not text. The page contains 0 text tokens but consumes hundreds of vision tokens depending on provider (see how many tokens in a PDF and how many tokens in an image).

Math and equations. LaTeX-formatted math is much denser per character than the visual equivalent. A page of equation-heavy physics can hit 800-1,200 tokens despite "only" 200-300 visible terms.

Non-English text. Languages with non-Latin scripts often tokenize less efficiently. Japanese and Chinese can run 2-3 tokens per character. Arabic and Hindi vary widely.

Markdown tables. A 10-row × 8-column markdown table eats roughly 250-400 tokens of just delimiter and alignment characters before the cell content. Tables in plain prose paragraphs are far more token-efficient.

Counting your specific page

The easiest path: paste a representative page into one of the model counters and read the exact number off:

Run two or three sample pages and average. That gives you a per-page rate you can multiply by the document's page count to estimate the whole document's tokens, usually within 5-10% of the truth.

Open the calculator →

Try this on every model

Try the live counter →