A token, in plain English
A token is a small chunk of text — sometimes a whole word, sometimes a word fragment, sometimes punctuation — that a language model reads and writes. Models do not process text character by character; they convert everything into tokens first. Tokens are both the memory unit (context windows are measured in tokens) and the billing unit (API prices are quoted per million tokens).
The rule of thumb used across the industry: one token is roughly 4 characters, or about three-quarters of an English word. That means 1,000 tokens is about 750 English words, and 1 million tokens is about 750,000 words — roughly 1,500 pages of a book.
What tokenization looks like
"Hello world" might become two tokens. A rare word like "tokenization" might split into token + ization. Dense text with lots of numbers, code, or non-English scripts tokenizes less efficiently — the same character count can cost noticeably more depending on content. This is why a long system prompt stuffed with boilerplate has a real dollar cost before the model generates a single word back.
What actually gets billed
Your bill is not "tokens in, tokens out" in the naive sense. Several token categories hit the meter, at different rates:
| Category | What it is | Typical rate |
|---|---|---|
| Input tokens | Your prompt: system message, documents, history, the user message | The headline rate — e.g. $2/1M on GPT-6.1 Sol |
| Output tokens | Everything the model generates, including hidden reasoning on thinking models | Usually 4–6× the input rate |
| Cached input tokens | Prompt prefixes the provider has seen before (see our prompt caching guide) | A steep discount — often ~10% of the input rate |
| Cache writes | The cost of writing a cache entry the first time | Above the input rate — e.g. 1.25× on GPT-6 |
| Image / audio tokens | Non-text inputs converted into token equivalents | Provider-specific; billed as input |
The hidden-reasoning line catches the most teams by surprise: on reasoning models, the model's chain-of-thought bills at output rates even though you never see it. A long-thinking answer can burn thousands of output tokens before the visible reply begins. Our reasoning vs fast models guide breaks that tradeoff down.
Token math: from units to dollars
Provider prices look like "$2 per million input tokens." To get the cost of one token, divide by a million: on GPT-6.1 Sol, one input token costs $0.000002 and one output token $0.00001. Individual tokens are fractions of a cent — volume is what creates the bill.
Example on GPT-6.1 Sol: 5,000 input + 800 output = 0.005 × $2 + 0.0008 × $10 = $0.018
Across current models, 1,000 tokens cost between roughly $0.0002 and $0.01 for input and $0.0012 and $0.05 for output, depending on the model. The full per-model spread lives on the JetAI pricing dashboard.
Why this matters before anything else
Tokens are the lever behind every saving on this site: caching re-prices input tokens, the batch API halves every token price, context windows cap how many you can send, and bill estimation is just arithmetic on your token counts. Learn the unit once; apply it everywhere.
Quick answers
How many words is 1 million tokens?
Roughly 750,000 English words, or about 1,500 book pages. The industry rule of thumb is one token per ~4 characters, or ~0.75 words, for English prose; code and non-English text tokenize less efficiently, so the same word count can cost noticeably more in tokens outside plain English text.
Why do output tokens cost more than input tokens?
Reading input is a parallel operation over your prompt, while generating output is sequential — each token requires a full forward pass through the model. That compute difference shows up as a 4x to 6x price markup on output across current models. Reasoning tokens also bill at output rates.
How many tokens is a typical email?
A short business email of ~200 words is about 270 tokens. A 10-page document runs roughly 5,000–7,000 tokens, and a long contract or research paper can pass 50,000. A 100K-line codebase runs into the millions — which is exactly why long-context pricing bands and caching discounts exist.
Do cached tokens count toward my bill?
Yes, but at a large discount. Cached input tokens typically bill at around 10% of the standard input rate — for example, GPT-6 Astra's cached input is $1.00 per million against $10.00 for fresh input. See our prompt caching guide for the break-even math.