LearnFoundations
Foundations

How we count tokens

What a token is, how input and output costs add up, and where to read the estimate on any listing before you buy.

Sigrix Editorial 6 MIN READ 533 WORDS

Almost every AI model — OpenAI, Anthropic, Google — bills by the token, not the word or the request. If you understand tokens, you can predict what a listing will cost before you ever run it.

§ 01 What is a token?

A token is roughly 4 characters of English text — about ¾ of a word. The sentence "Sigrix lists AI building blocks." is about 7 tokens. Models read and write in tokens, and providers bill per million tokens consumed.

The sentence "Sigrix lists AI building blocks." split into seven token chunks — even the word "Sigrix" is two tokens

Punctuation, spaces, and rare words split into more tokens; common words are a single token. Code and non-English text are usually denser, so the same character count costs more.

§ 02 Input + output = total cost

Every run has two sides. Input is what you send the model: the prompt, persona, attached skills, and any conversation history. Output is what the model writes back. Output tokens usually cost 3–5× more than input tokens, so a chatty model can be expensive even on a short prompt.

cost = (input tokens × input rate) + (output tokens × output rate)

That's why a one-line prompt that produces a 2,000-word essay can cost more than a long prompt that returns a single number.

Input — prompt, persona, skills, and conversation history — flows into the model, which returns output that costs 3–5× more per token; total cost is input tokens times the input rate plus output tokens times the output rate

§ 03 Why estimates vary by listing type

Each of the five building blocks sizes differently. The estimate on a listing reflects how that type tends to consume tokens:

Listing typeToken profile
Prompt / PersonaSmall input (a few hundred tokens). Output depends entirely on the task.
SkillThe SKILL.md file is loaded into context once, then reused at near-zero marginal cost.
AssistantMulti-turn, so input grows with conversation history. Expect 2–5× a single prompt.
AgentMulti-step runs accumulate tokens across each step plus tool calls. Heaviest by an order of magnitude.
ProductBundles all of the above, so the estimate is the sum of its parts.

§ 04 Where to find the estimate

Open any listing and check the Tokens tab. We show:

  • Estimated input tokens per run
  • Estimated output tokens per run
  • The model the seller tested on
  • A sample dollar cost at current provider rates

Estimates are seller-supplied averages. Actual usage depends on your inputs, the model you pick, and how long the conversation runs.

§ 05 Rule of thumb

When you compare two listings, look at the output estimate first — it's the bigger lever on cost. A listing with a large prompt but a tight, structured output is usually cheaper to run than a small prompt that rambles.

And if a listing doesn't show a token estimate yet, treat that as a question to ask the seller, not a reason to assume it's free.

0 Comments

Join the discussion. Be kind and stay on topic.

  1. No comments yet — be the first to share what you think.
Written by
Sigrix Editorial
Editorial · Marketplace & foundations

We write the playbooks we wish we'd had—field notes from building Sigrix and what the best creators do differently. Reach us at hello@sigrix.io.

One practical essay, every other Tuesday.

Patterns, sample listings, and short essays from the Sigrix editorial team. One-click unsubscribe, no noise.