What is Token?
By Heemang Parmar · Updated July 2026 · Editorial policy
A token is the basic unit of text that AI models read and generate, roughly four characters or three-quarters of an English word; model pricing, context windows, and generation speed are all measured in tokens.
The word 'understanding' might be one token or three depending on the model's tokenizer. As a working rule for English, a token averages about four characters, so 1,000 tokens is on the order of 750 words. Common words map to single tokens, while rare words, names, and punctuation get split into pieces.
Tokens matter to product builders because everything is denominated in them: API pricing is per million input and output tokens, context windows are token limits, and generation speed is measured in tokens per second. When an AI feature is slow or expensive, the first diagnostic is almost always how many tokens each request consumes.
The common confusion is equating tokens with words. Code, non-English text, and unusual formatting often tokenize less efficiently, so the same apparent length of content can cost multiples more. Providers ship token counters, and estimating a feature's per-request token budget early prevents invoice surprises later.
Why does token matter?
Tokens matter because they are the unit of account for the entire AI economy. Every model bill, rate limit, context window, and latency figure is denominated in tokens: a typical English word is about 1.3 tokens, and output tokens usually cost several times more than input tokens. A founder who cannot estimate a feature's token consumption cannot estimate its unit economics.
Token awareness separates profitable AI features from money losers. A chat feature that resends the full conversation history grows its input token count on every turn, so costs compound with engagement rather than staying flat. Techniques like prompt caching, history summarization, and context trimming can cut bills substantially without changing what users see.
How does token work?
- 1Split the text: A tokenizer breaks input into subword chunks drawn from a fixed vocabulary, typically tens of thousands of distinct tokens.
- 2Map to numbers: Each token becomes an ID and then a numeric vector that the model can actually compute with.
- 3Count against limits: Input and output tokens are tallied for billing, rate limits, and the context window ceiling on every request.
- 4Generate token by token: The model produces its response one token at a time, which is why long outputs take longer and cost more.
Token vs word vs character: what do AI models actually count?
| Unit | Rough size | What it is used for |
|---|---|---|
| Token | About 4 characters or 0.75 English words | Model pricing, context limits, and generation speed |
| Word | 1 to 1.3 tokens in typical English | Human-facing length estimates, not model accounting |
| Character | Roughly a quarter of a token | Storage and display, irrelevant to model billing |
How is token used in practice?
BYOK cost control
ProductOS lets you bring your own Anthropic, OpenAI, or Google API keys, so token usage is billed at provider rates directly on your own account with full visibility. Platform credits are the alternative if you would rather not manage provider keys yourself.
Routing to fit the task
Multi-provider model routing sends each job to an appropriate model rather than running everything on the most expensive one. Simpler tasks can run on cheaper models, which directly lowers the token cost per unit of work.
Shared context, spent once
ProductOS agents share one project context across research, PRD, design, code, and deployment. You describe the idea once instead of re-pasting the same background into every stage.
Frequently asked questions
How many tokens is a page of text?
A typical page of around 500 English words is roughly 650 to 700 tokens, using the rule of thumb that a word averages about 1.3 tokens. Code, tables, and non-English text usually tokenize less efficiently, so the same visual length can consume noticeably more. Provider token counters give exact figures.
Why do output tokens cost more than input tokens?
Generating text is more computationally expensive than reading it: each output token requires a full pass through the model, produced sequentially, while input tokens are processed largely in parallel. Providers price accordingly, with output commonly costing several times more per token, which is why verbose responses hit budgets hardest.
Do all models count tokens the same way?
No. Each model family uses its own tokenizer with its own vocabulary, so the same text yields different token counts on Claude, GPT, and Gemini. The differences are usually modest for English prose but grow for code and other languages, so always estimate with the target model's own token counter rather than a generic rule of thumb.
How can I reduce token costs?
Trim context to what the request actually needs, summarize long conversation histories instead of resending them, route simple tasks to smaller models, and use prompt caching for stable content like system prompts. Constraining output length helps too, since output tokens are the most expensive part of most requests.
Related terms
- Context windowA context window is the maximum amount of text, measured in tokens, that an AI model can process in a single request, covering the system prompt, conversation history, documents, and the model's own response.
- Large language model (LLM)A large language model (LLM) is an AI model trained on massive text datasets to predict and generate language, powering writing, coding, analysis, and reasoning tools through token-by-token text generation.
- Bring your own keys (BYOK)Bring your own keys (BYOK) is a usage model that routes an AI tool's requests through your own API keys from an LLM provider, so usage bills to your account and your data flows under your own provider agreement.
- EmbeddingAn embedding is a numeric vector representation of text, images, or other content that captures semantic meaning, letting software measure similarity between items and power semantic search, retrieval, and recommendations.