What does AI really cost to run? Token pricing in plain English
How AI is billed per token, what the main models cost today, and worked examples: what 1,000 emails summarised actually costs on each model.
Ali Hassan
· 6 min read
Quick answer
AI models used through an API charge per token, about 100 tokens for every 75 words, with separate prices for what you send in and what the model writes back. At November 2026 list prices, summarising 1,000 emails costs about $0.18 on a small model like Claude Haiku 5.5 and up to $17.50 on a top model like GPT-6 Astra. For most small businesses the AI usage is a few dollars a month; the setup costs more.
"What will the AI part cost us each month?"
I get asked this on almost every call, and the honest answer surprises people. For most small businesses, the AI itself is the cheapest line on the bill. Summarising a thousand emails can cost less than a coffee. But the price depends on which model you pick, and the difference between the cheapest and the most expensive is huge.
So let me explain how AI is actually billed, in plain English, with real numbers you can check.
You pay per token, not per question
When software talks to an AI model like Claude, GPT or Gemini through its API (the connection other software uses), you don't pay a monthly fee per person. You pay for tokens.
A token is a small chunk of text, usually part of a word. OpenAI's own rule of thumb is that one token is about four characters of English, and 100 tokens is about 75 words. Different models split text slightly differently. Anthropic, for example, says its newer models use a tokenizer that produces about 30% more tokens for the same text. So treat any estimate as a rough guide, and round up.
You're billed for two things:
- Input tokens: everything you send in. Your instructions, the email, any documents or company info it needs.
- Output tokens: everything the model writes back.
Output is always the more expensive of the two, usually about five times the input price.
You send text
Instructions, plus the email or document to work on.
It's counted as input tokens
Roughly 100 tokens for every 75 words.
The model writes a reply
Every word it writes is counted as output tokens.
You're billed for both
Input tokens × input price, plus output tokens × output price.
What the main models cost right now
Prices are quoted per million tokens. Here's a selection at the time of writing (November 2026), straight from each provider's pricing page:
| Model | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Claude Haiku 5.5 (small, fast) | $0.10 | $0.50 |
| GPT-6 Luna (small, fast) | $0.10 | $0.50 |
| Gemini 3.8 Flash (mid) | $0.75 | $3.75 |
| Claude Sonnet 5.5 (mid) | $2 | $10 |
| Claude Opus 5.5 (top) | $4 | $20 |
| GPT-6 Astra (top) | $10 | $50 |
These move often, in both directions. Google's pricing page, for example, already says the Gemini 3.8 Flash price rises to $1.50 input and $7.50 output from 1 January 2027. Always check the page before you commit.
Worked example: 1,000 emails summarised
Let's put real numbers on it. Say you want a short summary of every customer email, so your team can skim the inbox in minutes.
Here's the same job on each model in the table:
Cost to summarise 1,000 emails (1M input + 150k output tokens)
Claude Haiku 5.5
$0.18
GPT-6 Luna
$0.18
Gemini 3.8 Flash
$1.31
Claude Sonnet 5.5
$3.50
Claude Opus 5.5
$7.00
GPT-6 Astra
$17.50
The top model costs about 100 times more than the small ones for the same job. For summarising emails, a small model is usually good enough. You'd only pay for a bigger one if the small one keeps missing things, and you'd find that out by testing on your own emails.
A second example: drafting customer replies
Summaries are short. Replies need more context, because the AI has to know your prices, policies and tone.
That's why I said in Can AI actually answer my customers? that the AI usage is usually cents per conversation. The real cost is the setup and the checking, not the tokens.
The things that quietly push the bill up
The per-token price is only half the story. These are what I watch for when I estimate a project:
- Sending too much. If you paste a 40-page handbook into every request, you pay for those tokens every single time. Send only what the job needs.
- Long chats. In a back-and-forth conversation, each new message usually sends the earlier messages again, so the input grows with every turn.
- Retries and checks. A well-built system sometimes asks a second time, or has another call double-check the first. That's good practice, and it costs tokens.
- "Thinking" models. Some models reason step by step before answering, and that hidden reasoning is billed as output. Google's page says it plainly: output prices include thinking tokens.
- Picking the biggest model by default. As the chart shows, that alone can multiply the bill many times over.
API or a monthly subscription?
There's a second way to pay for AI: a per-person plan, like ChatGPT Business or Claude's Team plan. Claude's Team plan, for example, is $25 per seat a month on monthly billing at the time of writing.
My rule of thumb:
- People chatting with AI (writing, research, brainstorming): a per-seat plan is simpler and predictable.
- Software doing a job automatically (sorting emails, reading invoices, drafting replies): the API, paid per token. It's usually far cheaper for a defined task, and it can run without anyone opening a chat window.
Many businesses end up with both, and that's fine.
So what should you budget?
For a typical small business automation that uses AI on a few hundred or a few thousand items a month, the AI usage is usually a few dollars a month, sometimes less. Put that next to the build cost and tool subscriptions from what automation really costs, and you'll see the tokens rarely decide whether a project is worth it. Whether it saves real time does. I cover how to check that in how to measure the ROI of automation and AI.
Want a real estimate for your own task? Book a free 20-minute call. Tell me what you'd like AI to do and roughly how often, and I'll work out what it would cost to run, and whether a cheaper model would do the job.