Submit

Growth6 min read1,201 words

What Are LLM Tokens? How They Drive Your AI Costs

Tokens are the units an LLM uses to process text, and they are also the units you pay for. A token can be a character, part of a word, a whole word or punctuation, so token count is not word count and the difference shows up on your invoice.

ZeroRaised Team @zeroraised

Editorial title card reading "What Are LLM Tokens? How They Drive Your AI Costs" beside the orange ZeroRaised mark.

Key takeaways

  • A token is the unit a model uses to process text; it can be a character, part of a word, a whole word or punctuation.
  • In English, one token is roughly 4 characters or three-quarters of a word, and 100 tokens is about 75 words, but these are estimates.
  • Input, output, cached input and reasoning tokens are billed differently, and reasoning tokens are invisible but charged as output.
  • The same text can produce about 30 percent more tokens on Claude 4.7 and later models because of a newer tokenizer.
  • Price per token is not the same as cost per task, so test representative work before choosing a model.
On this page 10
  1. What is an LLM token?
  2. How many tokens are in a word?
  3. What are input, output, cached and reasoning tokens?
  4. How do you count tokens?
  5. Why do different models count the same text differently?
  6. Why price per token can mislead you
  7. How do you reduce token costs?
  8. What this means for bootstrapped founders
  9. Where ZeroRaised fits
  10. Sources

Tokens are the units a language model uses to read and write text, and they are the units you pay for. OpenAI describes them as "the units that OpenAI models use to process text", where a token can represent a character, part of a word, a whole word or punctuation (OpenAI).

TL;DR

  • A token is the model's unit of text, not a word; it can be a character, part of a word, a whole word or punctuation.
  • English rules of thumb are about 4 characters per token and about three-quarters of a word, and 100 tokens is about 75 words.
  • Input, output, cached input and reasoning tokens are priced differently, and reasoning tokens are billed as output.
  • The same text can produce roughly 30 percent more tokens on newer Claude tokenizers.
  • Judge models by cost per task, not price per token, then reduce input and cap output.

Disclosure: ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 3 Nov 2026; ZeroRaised has no commercial relationship with the companies named.

Token is the unit of text an LLM processes, and it can represent a character, part of a word, a whole word or punctuation (OpenAI).

What is an LLM token?#

A model does not read your prompt letter by letter or word by word. It breaks text into tokens using a tokenizer, then predicts the next token as it writes. OpenAI's wording is deliberately broad: a token can be a character, part of a word, a whole word or punctuation (OpenAI).

A short common word may be one token, while a long or unusual word may be several. The exact split depends on the model and the encoding, so the same sentence is not guaranteed to cost the same everywhere.

How many tokens are in a word?#

For English, OpenAI offers three rules of thumb: one token is about 4 characters, about three-quarters of a word, and 100 tokens is about 75 words (OpenAI). The help page labels these as estimates.

Use them for quick sizing, not for billing forecasts. Dense technical text, code, tables or non-English content can land far from the English rule, so count the actual text when a number matters.

What are input, output, cached and reasoning tokens?#

Providers split usage into categories because each one is priced differently (OpenAI):

  • Input tokens are the prompt: instructions, conversation history, documents, tool definitions and anything else you send.
  • Output tokens are what the model generates, including the visible answer.
  • Cached input tokens are input the provider has already processed and stored; pricing can differ, which is the basis of prompt caching.
  • Reasoning tokens are internal work on reasoning models. They are not visible but are billed as output tokens.

Because reasoning tokens are billed as output but never shown (OpenAI), you can plan for a short answer and still pay for a long think.

How do you count tokens?#

There are three practical routes on OpenAI: the Tokenizer, the tiktoken library using encoding_for_model, and an input-token counting API for the Responses API (OpenAI). Tools, schemas, images and files all affect the full count, so count the complete request rather than the user message alone.

Anthropic takes a similar approach. Count tokens before sending to manage rate limits and costs, make routing decisions and fit prompts. The count is an estimate, and you are not billed for system-added tokens (Anthropic).

Anthropic's token counting is free to use but subject to requests-per-minute rate limits by usage tier, separate from message creation limits (Anthropic).

Why do different models count the same text differently?#

Tokenizers change. Anthropic states that Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer, and the same input text produces approximately 30 percent more tokens than on earlier models (Anthropic).

A headline rate per million tokens is only comparable if you also know how many tokens your text becomes on that model. A 30 percent difference in token count moves the effective cost even when the quoted rate looks better.

This is also why you should not port a prompt budget from one provider to another without re-counting. The context window guide covers the related limits.

Why price per token can mislead you#

OpenAI's own advice is blunt: "A lower price per million tokens does not necessarily produce a lower total cost" (OpenAI). The recommendation is to test representative tasks.

The GDPx leaderboard makes the gap visible. It lists Claude Opus 5.5 at a 12.0% score with an average cost per task of $51.41, GPT-6 Luna at 5.1% and $0.92, and DeepSeek 4.1 Flash 1017 at 5.4% and $0.94 (GDPx leaderboard). The cheap models cost a tiny fraction per task, but they also score lower. The right choice depends on what a correct task is worth to you, not on the sticker price alone.

Run this on your own workload: pick ten real requests and compare total cost to complete each one, including retries. The LLM cost guide walks through the wider set of levers.

How do you reduce token costs?#

OpenAI's suggestions start with the input, then cap the output (OpenAI):

  1. Shorten the prompt. Remove polite filler, duplicated instructions and examples that no longer earn their place.
  2. Remove repeated context. Do not resend the same document on every turn if the model already has it.
  3. Split the job. A smaller task with a smaller prompt is often cheaper and easier to check.
  4. Summarise history. Replace long transcripts with a compact summary that keeps the decisions.
  5. Cap the output. Use max_completion_tokens or max_output_tokens so a runaway generation cannot set the bill.

Then look at caching for the stable prefix and routing for low-value calls, covered in the prompt caching guide and the cost guide.

What this means for bootstrapped founders#

Tokens are a unit cost, and unit costs decide gross margin. If you price a subscription without knowing the tokens a typical task consumes, a heavy user can quietly turn a profitable plan into a subsidy.

Measure three numbers: tokens per task, cost per task and tasks per customer per month. Then set limits or prices from those numbers rather than from the average. The AI pricing guide and the usage limits guide cover how to turn that into a product decision.

Do not optimise the wrong thing. Cutting a fraction of a cent from a weekly prompt is not worth the engineering time. A 30 percent token increase on a prompt that runs on every request is where the money is.

Related reading from this week: Runway and BoldDesk Swap Flat AI Plans for Credits. Price Like This and Larridin Raised $30M Because AI Bills Got Scary. Sell ROI, Not Seats.

Where ZeroRaised fits#

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have a product people can try, you can submit it here. The pricing page explains the current options.

For more practical analysis, join the ZeroRaised newsletter. No promises of traffic or results, just the useful parts.

Sources#

Frequently asked questions

What is a token in an LLM?

A token is a unit of text that a model processes. It can represent a character, part of a word, a whole word or punctuation. The model reads tokens and generates tokens, and providers usually bill per million input and output tokens.

How many tokens is a word?

There is no fixed answer because tokenizers differ. OpenAI's rule of thumb for English is that one token is about 4 characters or about three-quarters of a word, and that 100 tokens is about 75 words. Treat those figures as estimates and measure your own text.

Are reasoning tokens billed?

Yes, on the models that expose them. OpenAI describes reasoning tokens as not visible in the output but billed as output tokens. That means a long internal reasoning pass can raise the bill without appearing in the answer.

Does a cheaper model always cost less?

No. A lower price per token does not necessarily produce a lower total cost, because a cheaper model may need more tokens or more attempts to finish the same task. Compare cost per completed task, not the headline rate.

Does the same prompt cost the same on every model?

No. Tokenizers differ between providers and between model generations. Anthropic says Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer, and the same input text produces approximately 30 percent more tokens than on earlier models.