Charts · 30

A context window, to scale

A context window is how much a model can read at once, and it is measured in tokens, not pages. Windows grew from 8,192 tokens to a million in under a year, which sounds like infinity until you put something real in it. These are my own things, counted with a real tokeniser: everything I have published on this blog fits in one window with room to spare, and Lunary, at 11.4 million tokens, would need 12 of the largest.

Context window
Axis

A 1M window is 1,000,000 tokens: about 3,030 pages. Lunary would need 12 of them, end to end.

  1. A page250 words, estimated330fits 3,030×
  2. A short story7,500 words, estimated9,892fits 101×
  3. A novel40,000 words, estimated53kfits 18×
  4. My blogall 66 posts, measured113kfits 8×
  5. sammii.dev325 files, measured430kfits 2×
  6. PostReady330 files, measured934kfits 1×
  7. Spellcast824 files, measured1.1M1.1 windows
  8. Lunary5,337 files, measured11.4M11.4 windows

code · content, docs and data in the same repo · estimated from a word count. The dashed line is the window. On the to-scale axis, anything under about 100k tokens is a sliver: switch to log to read the small ones.

1M · 1,000,000 tokens

Gemini 1.5 Pro (2024) and GPT-4.1 (2025)

Hover a bar for what was counted. The four codebases are mine, counted file by file with a real tokeniser; only the totals are published here.

Every count, o200k_base and cl100k_base, counted 19 October 2026
WhatFilesCodeTotalcl100k
PostReady330843,855933,723930,315
sammii.dev325261,250430,231428,296
Spellcast8241,014,4511,123,9351,114,546
Lunary5,3379,171,66711,377,43611,343,051
My blog66n/a112,552113,260

Notes

Counted on 19 October 2026 with OpenAI’s tiktoken 0.12.0, encoding o200k_base (the tokeniser of GPT-4o, GPT-4.1, o3 and GPT-5) with cl100k_base (GPT-4) alongside; the two agree within 1% on all four codebases. Claude and Gemini use their own tokenisers, so the same files are a somewhat different number of tokens to them, but the order of magnitude does not change. Each codebase is every file git tracks at the tip of its default branch, minus dependency and build folders, lock files, generated files, minified bundles, binaries, images and fonts. Code is source in a programming language; everything else is content, documentation and data. Only totals leave the counting script; no file names or code are published.

The page, short story and novel are estimates: a word count by definition times 1.319 tokens per word, measured on the prose of my 66 blog posts with code blocks removed (70,112 words, 92,473 tokens), close to OpenAI’s own rule of thumb of 100 tokens for 75 words. A page is the publishing industry’s 250 words for a double-spaced manuscript page. The short story and novel are the boundaries in the SFWA Nebula Awards rules: a short story is under 7,500 words, a novel is 40,000 or more.

Windows: GPT-4 launched with 8,192 tokens and a 32,768-token variant (OpenAI, GPT-4, March 2023); GPT-4 Turbo had 128,000 (OpenAI DevDay, November 2023); Claude 2.1 had 200,000 (Anthropic, November 2023); Gemini 1.5 Pro offered 1,000,000 (Google, February 2024), and GPT-4.1 1,047,576 (OpenAI, April 2025). A window holds the instructions, the conversation and the reply as well as whatever you paste, so the usable space is always less than the number. Scripts and method: scripts/labs in this site’s repository.