A field guide to AI pricing
Every AI API bills you per million tokens. Nobody tells you how much that actually is. Here's the answer — in novels, photos, podcasts, and dollars.
The scale
Sixteen ways to picture the same number. Each estimate shows the assumption it's built on.
1 block = 750 words
≈ 750,000 words
English words
On average one token is about ¾ of an English word, so a million tokens works out to roughly three quarters of a million words.
basis: 1 token ≈ 0.75 words
1 block = 1 page
≈ 1,500 pages
single-spaced pages
At about 500 words per single-spaced page, 750,000 words fills roughly 1,500 pages — a stack of paper about 15 cm (6 in) tall.
basis: 500 words per page
1 spine = 1 novel
≈ 8 novels
average novels
A typical novel runs about 90,000 words (~120,000 tokens), so a million tokens holds roughly eight of them.
basis: 90,000 words per novel
the 7 books — filled part = 1M tokens
≈ 69% of Harry Potter
of the complete 7-book series
The full Harry Potter series is 1,084,170 words — about 1.45 million tokens. A million tokens gets you from Privet Drive to somewhere deep in Order of the Phoenix.
basis: Series total 1,084,170 words
2 copies — filled part = 1M tokens
≈ 1.3 × War and Peace
copies of War and Peace
Tolstoy’s epic is about 587,000 words in English translation (~780,000 tokens). A million tokens covers it with room for a second read of the first third.
basis: 587,000 words per copy
1 block = 100 lines of code
≈ 100,000 lines of code
lines of source code
Code averages roughly 10 tokens per line, so a million tokens is about 100,000 lines — nearly twice the original Doom engine (~57,000 lines).
basis: ~10 tokens per line
1 tile = 1 image
≈ 1,000 images
detailed images analyzed
Vision models spend roughly 800–1,600 tokens to read one detailed image, so a million tokens lets a model look at about a thousand photos or screenshots.
basis: ~1,000 tokens per image
1 block = 1 PDF page
≈ 650 PDF pages
visually-processed PDF pages
When a model reads a PDF page as both text and image (preserving layout, tables and figures), each page costs ~1,500 tokens. Text-only extraction is cheaper: ~2,000 pages per million tokens.
basis: ~1,500 tokens per page with layout
1 block = 25 messages
≈ 37,000 chat messages
chat messages
An average chat message is about 20 words (~27 tokens). A million tokens is over three years of texting at 30 messages a day.
basis: ~27 tokens per message
1 bar = 1 hour of talking
≈ 83 hours of speech
hours of transcribed speech
People speak about 150 words per minute (~200 tokens/min), so a million tokens transcribes about 83 hours — three and a half days of non-stop talking.
basis: 150 words per minute
1 block = 1 episode
≈ 110 podcast episodes
45-minute episodes
At ~9,000 tokens per 45-minute episode transcript, a million tokens holds about 110 episodes — a full year of a weekly show, twice over.
basis: 45 min per episode, 150 wpm
1 block = 10 posts
≈ 14,000 max-length posts
280-character posts
A maxed-out 280-character post is about 70 tokens, so a million tokens is roughly 14,000 of them — about 10 years of posting four times a day.
basis: 280 chars ≈ 70 tokens
1 block = 10 emails
≈ 10,000 work emails
typical work emails
The average work email body runs about 75 words (~100 tokens). A million tokens is ten thousand of them — several years of a busy inbox.
basis: ~100 tokens per email
1 spine = 1 screenplay
≈ 37 movie screenplays
feature-film screenplays
A feature screenplay is around 20,000 words (~27,000 tokens), so a million tokens holds about 37 movies’ worth of scripts.
basis: 20,000 words per screenplay
1 block = 1 article
≈ 1,100 Wikipedia articles
average Wikipedia articles
The average Wikipedia article is about 660 words (~880 tokens), so a million tokens is over a thousand articles of general knowledge.
basis: ~660 words per article
1 bar = 2 songs
≈ 2,100 songs
songs’ worth of lyrics
Song lyrics average ~350 words (~470 tokens), so a million tokens is the lyrics to about 2,100 songs — over five days of continuous music.
basis: ~350 words per song
The price
Pick a budget and see how many novels' worth of text each model writes (or reads) for it. Prices updated 2026-07-03.
One spine = one novel written for $10. Same money, 60× difference between the cheapest and priciest model. Filtering the scale section above switches the unit here too.
| Model | Provider | Input / 1M | Output / 1M | Context |
|---|---|---|---|---|
| GPT-5.5 Pro | OpenAI | $30 | $180 | 400K |
| Claude Fable 5 | Anthropic | $10 | $50 | 1M |
| GPT-5.5 | OpenAI | $5 | $30 | 400K |
| Claude Opus 4.8 | Anthropic | $5 | $25 | 1M |
| Claude Sonnet 5 | Anthropic | $2 | $10 | 1M |
| Gemini 3.1 Pro | $2 | $12 | 1M | |
| Gemini 3.5 Flash | $1.50 | $9 | 1M | |
| Claude Haiku 4.5 | Anthropic | $1 | $5 | 200K |
| Gemini 3 Flash | $0.50 | $3 | 1M |
GPT-5.5: Long-context (>272K input) priced at 2× input / 1.5× output. Claude Sonnet 5: Introductory pricing — list price is $3 / $15 from Sep 1, 2026. Gemini 3.1 Pro: Prices double for prompts above 200K tokens ($4 / $18).
The calculator
Pick a model, set a budget, and see what it translates to.
1,000,000tokens
The fine print
A token is the unit AI models read and write text in — usually a word, part of a word, or a punctuation mark. "Cat" is one token, but "tokenizer" splits into pieces like "token" + "izer". In typical English, one token is about ¾ of a word, so 1,000 tokens ≈ 750 words.
Generating text is much more computationally expensive than reading it. Every provider reflects that: output tokens typically cost 5–6× more than input tokens. Your real bill depends on your mix — a summarization app sends lots of input and gets little output, while a writing app is the opposite.
No. Each model family uses its own tokenizer, so the same text can differ by 30% or more in token count between providers — and even between model generations from the same provider. That means a per-million-token price is only comparable once you account for how each tokenizer counts your text.
Every major provider offers a token-counting endpoint or tokenizer tool (Anthropic has a count-tokens API, OpenAI and Google publish tokenizers). Rough planning is easier: divide your word count by 0.75, or your character count by 4.
Anywhere from $0.50 (Gemini 3 Flash input) to $180 (GPT-5.5 Pro output) as of July 2026 — a 360× spread. That is exactly why it pays to know what a million tokens is before choosing a model.