A token is the small chunk of text an AI model reads and generates, and it's the unit almost every AI tool bills you on. One token is roughly four characters, though that number moves around depending on the word (Zapier, Harry Guinness, 2026-07-07; corroborated by OpenAI's Help Center).
The short answer: a token is a chunk of text (and the thing you pay for)
Quick disambiguation before we go any further: this is not the crypto “AI token” you'd buy on an exchange. An AI-model token is a unit of text a model reads and bills you on. A crypto AI token is a tradeable blockchain asset, usually just branded around an AI theme to sound current. Same word, nothing else in common.
One token is roughly four characters of text, the unit you pay for. It is an estimating guideline, not an exact formula, since the count moves around by word and language.
By the end of this piece, you'll be able to open any AI pricing page, see the word “token,” and actually work out what something is going to cost before you run it. If you're still fuzzy on what an AI agent actually is before we get into billing units, that's the place to start first.
How text becomes tokens (the tokenizer, in plain English)
Every model has a tokenizer, the part that “splits text into tokens, the small chunks an AI model reads and generates” (Zapier, 2026-07-07). Each token gets a unique ID, and every model tokenizes slightly differently. That's one reason the same prompt can cost different amounts on different platforms, even when it looks identical to you.
Here's a concrete one: the URL https://www.harryguinness.com/blog/new-site-announcement breaks into 14 tokens per OpenAI's public tokenizer tool (via Zapier, 2026-07-07). Short words like “the” usually get their own token. Longer or rarer words split into two or three pieces.
Whitespace and capitalization matter too. “and” and “ and” (with a leading space) count as different tokens. OpenAI's Help Center shows the same word gets different token IDs depending on capitalization and position: “ red,” “ Red,” and “Red” at the start of a sentence are three separate tokens. The model doesn't see words the way you do. It sees these fragments.
One sentence on what happens next, and this is as deep as this needs to go: after splitting the text into tokens, the model converts each one into numbers so it can do math on them.
This is also why “count the characters” or “count the words” doesn't work as a formula. There's no universal conversion rate (Zapier, 2026-07-07). Numbers, code, and other languages tokenize differently than plain English prose. The roughly-4-characters-per-token rule is for sanity-checking a cost before you commit budget, not a formula you can trust down to the decimal.
Input tokens vs output tokens (why the same task can cost more)
Input tokens are what the model reads: your prompt, a pasted email, any context you feed it. Output tokens are what it generates back (Zapier, 2026-07-07).
That's a real lever you can pull. A workflow that reads a lot but writes a little, like classifying an incoming email as “urgent ” or “spam,” is cheap. One that writes a lot, like drafting a full 1,000-word reply for every lead, burns output tokens fast. If a workflow that seemed simple on paper has been chewing through your credits faster than you expected, this is usually why. Check which side of that split it leans on before you assume something's broken.
Worked estimate, not a measurement: a roughly 200-word email is about 250 input tokens, using the roughly-4-characters-per-token guideline. At Haiku 4.5's rate of $1 per million input tokens, reading that email costs a fraction of a cent. Reasoning models add a wrinkle here. Cranked-up reasoning effort can generate a hidden chain of thought before the model writes a single visible word, sometimes “tens of thousands of tokens” of invisible thinking (Zapier, 2026-07-07), and all of it still billed as output.
Context windows: how much the model can hold at once
A context window is “the maximum number of tokens a model can process in one go” (Zapier, 2026-07-07), counting both what you send in and what the model sends back.
This is where non-coders get bitten without realizing it. Paste a long document plus a long chat history and you can blow past the window entirely. Older models had it rough: GPT-3.5 ran on a 4,096-token window, “which is why they kept forgetting things mid-conversation” (Zapier, 2026-07-07). If you ever used one of those early tools and it just... forgot what you'd told it three messages ago, that's the reason.
Modern frontier models are far bigger. As of 2026-07-01, Claude's documented context window sits at 1 million tokens (Zapier's Claude AI guide), large enough to “hold entire codebases.” TechCrunch confirms current-generation frontier models “typically offer 1 million tokens or more” as of 2026-06-30, though specs drift fast, so check the vendor's page before building around an exact number.
One more nuance worth knowing: as a conversation nears the context limit, most models auto-summarize it and carry the summary forward rather than cutting off abruptly (Zapier, 2026-07-07). Convenient, sure, but the model is then working from a digest, not the original, and specific details can quietly get lost in that compression.
Operator takeaway: a bigger window is not a free pass. You still pay per token inside it, regardless of size.
How tokens turn into your actual bill (the table the others skip)
Here's the confusion almost nobody addresses: most non-coders never see the word “token” on their invoice. You see “tasks” or “credits” instead, and that hides the token underneath completely.
Raw API pricing is quoted per million tokens, priced separately for input and output. As of 2026-06-30 (TechCrunch), the current lineup: Claude Fable 5 at $10/$50 per million tokens, Opus 4.8 at $5/$25, Sonnet 5 at $2/$10 (rising to $3/$15 after 2026-08-31), and Haiku 4.5 at $1/$5. All four share a 1-million-token context window. Sonnet 5 is worth naming on its own: TechCrunch calls it cheaper than Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, while still pricier than Gemini 3.5 Flash. Prices drift constantly, so treat this as a snapshot and check the vendor's page before committing budget.
Raw per-token pricing is only one system you'll run into, though. Here's how the pricing units you actually see on automation platforms map back to tokens:
| Pricing unit | What it measures | Example tool | What to watch |
|---|---|---|---|
| Raw tokens (input/output) | Text read vs. generated, billed separately, per million | Anthropic and OpenAI APIs | Output costs more than input; long replies add up fast |
| Tasks | One action step per workflow run, regardless of token count inside it | Zapier | Triggers are free; a single AI step can hide a lot of token usage in one “task” |
| Credits | A weighted unit per AI call | Gumloop | Standard call ~2 credits, advanced ~20, expert 30+; triggers cost credits too (unlike Zapier); no rollover |
| Compute-time | Runtime duration times memory, not step count | Pipedream | ~1 credit is about 30 seconds of compute at 256MB memory; a slow call burns more than a fast one for the same job |
(Sourced to Zapier's Gumloop and Pipedream pricing breakdowns, 2026-06-08 and 2026-07-06.)
Whichever platform you're on, tokens are the real cost driver underneath. Zapier just wraps them in a flat per-task price, Gumloop weights them by call complexity, and Pipedream bills the clock time instead. Different wrapping, same engine. For more on this, see how automation pricing actually works.
Once you've got the pricing units straight, the next question is which model to run behind them. We cover that in which AI model to use for automation.
If pricing pages like this make your eyes glaze over, that's exactly what the newsletter untangles, plain English, no hype, sent occasionally rather than daily.
How to keep your token bill sane (practical, no code)
Match the job to the model. Send lightweight, high-volume tasks to a cheaper model, and save expensive reasoning models for jobs that actually need them (Zapier, 2026-07-07). Haiku or Sonnet-tier models handle most classification and short-reply work just fine. Save Opus or Fable-tier for genuinely hard, multi-step reasoning, not for everything by default.
Trim your input. Don't paste a whole document when a paragraph would do. Watch your system prompts too, a long instruction block repeating on every run compounds across every single execution.
Watch the output side hardest, since it costs more per token than input. Cap reply length where you can, and ask for structured or short outputs instead of open-ended prose.
Reality check: on Zapier's own AutomationBench, a proprietary benchmark, not an industry-standard eval, the top-scoring model completed only about 17.4% of real multi-step workflow patterns unaided (Zapier, 2026-06-10). A bigger reasoning model does not guarantee success. Don't pay a premium for a job a cheap model already handles, and if you're not sure whether you need an agent at all, that's worth checking first.
Frequently asked questions
How many characters is one token? Roughly four characters, though it varies by word and language (Zapier, 2026-07-07; OpenAI Help Center). Treat it as an estimating guideline, not an exact formula.
What is the difference between input and output tokens? Input tokens are what the model reads, output tokens are what it generates, and output usually costs more because generating text takes more work than reading it (Zapier, 2026-07-07).
Is an AI token the same as a crypto token?No. An AI-model token is a unit of text a model reads and bills you on. A crypto “AI token” is a tradeable blockchain asset branded around AI. They share a name and nothing else.
How do I count tokens without coding?Use the roughly-4-characters-per-token rule of thumb, or check a vendor's public tokenizer tool for an exact count on a specific piece of text. No code required for a rough number.
Why do longer prompts cost more? More text means more input tokens, and a long system prompt repeating on every run adds up across every single execution of a workflow.
What is a context window?The maximum number of tokens a model can process in one go, counting both what you send in and what it sends back (Zapier, 2026-07-07). Current frontier models document windows up to 1 million tokens as of mid-2026, though the figure drifts and is worth checking on the vendor's page.
The one-sentence version (and where to go next)
A token is the chunk of text an AI reads and writes, and it's the unit you pay for, so watch your output and match the model to the job.
If this kind of plain-English breakdown is useful, the newslettercovers this stuff regularly, no hype, just what's worth knowing. From here, it's worth understanding what an AI agent actually is, or going straight to which AI model to use for automation now that you know what's actually driving the price.