Token

The chunk of text a language model reads or writes; roughly four characters of English per token.

A token is the unit of text that language models actually process. Models do not see letters or words: they see tokens produced by a tokenizer, and every price, limit and context size in AI infrastructure is denominated in them. In English, a token averages about four characters, so 1,000 tokens is roughly 750 words.

Why it matters to an indie builder: tokens are the currency of AI features. Context windows, API pricing, embedding storage and rate limits are all token-based. If you can estimate tokens, you can estimate cost before you write a line of integration code.

A concrete example: your app sends a system prompt (800 tokens), the last six chat messages (2,200 tokens) and a user question (150 tokens). Input is 3,150 tokens. The model answers with 350 tokens, which you also pay for, usually at a higher rate. If input costs $0.30 per million tokens and output $1.50 per million, this exchange costs about $0.0014. Ten thousand such exchanges a month is $14: know this number before you promise the feature.

Common mistakes: assuming one word equals one token (code and non-English text tokenize worse; Spanish runs about 25 percent more tokens than English for the same content), ignoring output pricing, and blowing the context window mid-conversation, which forces silent truncation or errors. Practical habits: count tokens for real with a tokenizer before quoting costs, cache repeated prompts (Anthropic and OpenAI both discount cached input), and log token usage per feature with Langfuse or Helicone so the bill never surprises you.

One practical habit: build with a token budget from day one. A simple logger that records tokens per request and per feature turns an opaque monthly bill into a chart you can act on. The llm-ops tools listed on IndieBoot (Langfuse, Helicone, Portkey) do this out of the box.