Skip to content
AuraMatrix
← All models

Token Price Index · A guide

What does a token price actually mean?

AI models are sold by the token, at prices like “$2 per million”. Here is what that buys: what it costs to ask a question, read a report or run an assistant for an hour, worked out from today's official list prices of 23 models.

Prices as of 2026-11-15 22:15 UTC202 price records, each with its sourceFigures are estimates from list prices

$3.00 to read, $15.00 to write, per million tokensSources →

1 · The unit

A token is a piece of a word

Models do not count words or characters. They count tokens: short pieces of text, about four letters each in English, about one character each in Chinese. You pay for every token a model reads and every token it writes.

About 13 tokens. Reading them costs under 0.01¢ on Kimi K3. The split shown is an illustration; each model cuts text its own way.

One million tokens is about750,000English words
or1,500pages
or8novels, cover to cover

2 · Two prices

Reading is cheap. Writing costs more.

Every model has two prices: one for the text it reads (input) and one for the text it writes (output). On Kimi K3, writing costs 5 times as much as reading. Having a model read a long report is cheap. Having it write a long one is not.

One dollar reads about500pages (2.8 novels)
One dollar writes about100pages (0.56 novels)

3 · Real jobs

What a job costs

A price per million tokens means little until it is attached to a piece of work. Pick one.

A short question, a few paragraphs back.

0.81¢

on Kimi K3, at today's list price

  • Reading 200 tokens 0.06¢
  • Writing 500 tokens 0.75¢

Do it 20 a day and a month comes to about $4.86.

The same job on every modelCheapest first. Select a row to switch model.

4 · The levers

What makes a bill grow, and what shrinks it

A single question costs next to nothing. Bills come from volume: programs that re-read the same long context on every step, and models that “think” before they answer. Take the coding assistant: in one hour it reads 1.5M tokens and writes 60K. Pull the levers.

How it runs
Thinking before answering

$5.40

  • Reading $4.50
  • Writing 90¢

5 · Where you buy

Same model, different shop, different price

A model is sold by its maker, by cloud platforms and by marketplaces. They do not all charge the same. This is Kimi K3 on every channel in the index, price to read one million tokens.

  1. Kimi APIthe maker’s own price$3.00reference
  2. Amazon Bedrock, global cross-regioncloud platform$3.00same
  3. Amazon Bedrock, US East (N. Virginia)cloud platform$3.30+10%

Kimi API, China site ¥20 to read, ¥100 to write, per million tokens. China-site prices are shown in CNY as published. They are not converted, because an exchange rate moves every day and a list price does not.

6 · Over time

Prices move

List prices are not fixed. Makers cut them, promotions end, and a marketplace's price follows whichever provider it routes to. The index re-reads every source daily and keeps every reading. Since 2026-11-03 it has recorded 27 price changes.

DeepSeek V4 Pro on OpenRouter: price to read one million tokens
$0$1$23 Nov9 Nov13 Nov$0.9

$0.9 to read, $1.5 to write read 13 Nov, 22:15 UTC

Read this too

What these figures are, and are not

  • List prices, not bills. Every figure is calculated from the price each channel publishes, re-read daily. No requests were sent and no invoices were read.
  • Job sizes are illustrations. A “30-page report” is taken as 20,000 tokens. Your documents, your prompts and your model's way of cutting text will differ.
  • Token counts are rules of thumb. About 0.75 English words per token; other languages and other models count differently.
  • Some rules are not public. Where a price page does not say how something is billed, the index says “unknown” and its estimates return a range, not a guess.

See Kimi K3, with sourcesAll 24 modelsHow prices are checkedAsk your own AI assistant