Decisions · Notes

What is a token in AI? How many words it is, and what it costs

One token is about four characters of English, so 1,000 tokens is roughly 750 words. What tokens cost on the main AI APIs, and why output costs more.

Alan Lui · 19 November 2026

A token is the unit AI models read and write text in: a word, part of a word or a punctuation mark. As a rough guide for English, one token is about four characters, so 1,000 tokens is roughly 750 words. AI APIs charge by the token, usually quoted per million tokens, and they charge more for the tokens the model writes (output) than for the ones you send it (input).

How many words is a token?

Perplexity's API documentation puts it this way: “1 token is about 4 characters of English text. The exact count varies with language and content.” At an average of about five characters per English word including the space, that works out at roughly three-quarters of a word per token, or about 750 words per 1,000 tokens. Other languages usually take more tokens for the same meaning, and code and numbers vary.

What does a token cost?

Prices are quoted per million tokens (1M). From the providers' own pricing pages, checked 18 November 2026:

ModelInput, per 1M tokensOutput, per 1M tokens
Claude Opus 5.5 (Anthropic)$4$20
Claude Sonnet 5.5$2$10
Claude Haiku 5.5 (prompts up to 100K tokens)$0.10$0.50
DeepSeek V4.1 Flash, peak hours$0.30$1.20
DeepSeek V4 Pro, peak hours$1.32$3.96

So a million tokens of input, about 750,000 words, costs from $0.10 to $4 on these models. Every provider's current rates, side by side, are on What do AI model APIs cost?.

Why are output tokens more expensive?

On every model above, output costs several times input: five times on the Claude models, three to four times on DeepSeek. Because providers price what the model writes higher, a job that produces long answers costs far more than one that reads a lot and answers briefly. When you estimate a bill, count the two separately.

What are cached input tokens?

If you send the same long instructions or documents again and again, providers can cache them and charge much less the second time. Anthropic charges $0.20 per million cached tokens on Opus 5.5 against $4 for fresh input; DeepSeek charges $0.006 per million on V4.1 Flash at peak against $0.30. Writing to the cache can cost a little more than normal input (Anthropic: 1.25 times for a five-minute cache).

Do ChatGPT or Claude subscriptions charge by the token?

No. A subscription such as Claude Pro or ChatGPT Plus is a flat monthly price with usage limits; tokens are how the API is billed, for developers and apps. When you hit a subscription's limit, some plans let you keep going at API rates (see Claude usage limits explained).

Sources: Perplexity API docs (tokens and characters) · Anthropic API pricing · DeepSeek API pricing. Read on 18 November 2026. The words-per-token figure is our arithmetic from the four-characters guide, not a provider's number.

The page this is about: What does an AI model API really cost to run? — every figure sourced to the vendor's own page, with the date it was checked.