Token-Based Pricing: How It Works, What It Costs and When to Use It

Years ago, I was the government national sales manager for a telecommunications company. We sold on a usage-based model: the more calls you made, the more you paid.

One month, a government customer made a crazy number of conference calls during an election. Their bill went through the roof. We celebrated.

The next month, their spending dropped off a cliff. We panicked.

I tell that story in the book to show the "feast or famine" risk of usage-based pricing. I'm telling it again because AI has brought the phone bill back. Only this time the minutes are called tokens, and the businesses getting the surprise bills are the ones building AI products.

If you sell anything with AI inside it, you need to understand tokens. You don't necessarily need to charge for them. Let me show you the difference.

What is a token in AI pricing?

A token is a chunk of text that a language model reads or writes. It might be a whole short word, part of a longer word, or a punctuation mark.

OpenAI's rule of thumb is that one token is about four characters of English text, or roughly three-quarters of a word. So 1,000 tokens is about 750 words, and a 300-page book is a few hundred thousand tokens.

There are four kinds of tokens on a typical bill:

• Input tokens: everything you send the model. Your instructions, the customer's question, any documents or past conversation.
• Output tokens: everything the model writes back. These cost more.
• Reasoning tokens: the "thinking" a reasoning model does before answering. These are usually billed as output.
• Cached input tokens: input the vendor has seen recently and stored, billed at a steep discount.

Here's the catch most people miss. A model has no memory between calls. If your chatbot is ten messages into a conversation, the eleventh call sends all ten previous messages again. You pay for them again.

How does token-based pricing work?

Model vendors quote a price per million tokens, often written as "per MTok," with separate rates for input and output.

Here's the ladder as it stands today. Prices checked 27 October 2026; they change often, so see the AI Pricing Index for the live tracker.

Three things stand out.

Output costs more than input. Usually five times more at OpenAI, Anthropic and Google, three times at SpaceXAI and Mistral.

The spread is 100x. A small model costs a hundredth of a frontier one. Choosing the right model for each job is the biggest cost lever you have.

Some prices go up. Google has scheduled Gemini 4 Argon and Gemini 3.8 Flash to double after their launch periods. If you build on them, cost your product at the higher price.

For every model side by side, with the discounts and surcharges, see the LLM API pricing comparison.

How is token cost calculated?

Here's the formula (the glossary has a shorter cost per token explainer). You'll use it constantly.

Cost per task = (input tokens × input price per 1M + output tokens × output price per 1M) ÷ 1,000,000

Let's run it on a real job.

Same product, same customers. The monthly bill ranges from $166 to under $4 depending on two design choices.

Now compare that with what the customer values. Fin charges $0.99 per resolved conversation. HubSpot charges $0.50. Your cost is measured in fractions of a cent, and your customer's value is measured in dollars. Keep that gap in mind. It's the whole argument for not charging your customers per token.

Why do AI agents cost so much more per task?

A chat question is one call. An agent is a loop.

It plans, calls a tool, reads the result, thinks, calls another tool, and so on. Every step resends the growing context. Reasoning models add their own thinking tokens on top.

Take a 10-step research agent on Claude Sonnet 5.5 whose context grows from 5,000 to 50,000 tokens. That's about 275,000 input tokens and 20,000 output tokens: $0.55 + $0.20 = $0.75 per run. A one-line chat question on the same model (1,000 tokens in, 300 out) costs $0.005.

The agent costs 150 times more. And to the customer, both feel like "asking the AI to do something."

That's exactly why the developer tools repriced in 2025 and 2026. GitHub (see how Copilot moved to credits) said a short chat question and a multi-hour agent session had cost the same under its old model, which was "no longer sustainable." As one analysis put it, "a lower listed price can still mean a higher total cost." For the pricing side of agents, see AI agent pricing.

How do you cut a token bill?

There are five levers, and the vendors publish all of them:

Watch for the surcharges too. Google charges about 1.8x for Priority processing and more for prompts over 200,000 tokens on its Pro models. OpenAI adds 10% for data residency. Anthropic's cache writes cost 1.25x input for five minutes and 2x for an hour, so caching only pays when the cached text is reused.

Every one of these feeds straight into your gross margin. See AI gross margins and inference costs for how they add up.

Should you charge your customers per token?

Sometimes. But less often than you'd think.

When tokens work:

• Your buyers are developers. They expect to see the meter and they can forecast it. Every model vendor sells this way.
• You resell inference openly. Vercel's v0 publishes its own token rates by tier, from v0 Mini at $0.20 in and $1.20 out to v0 Max Fast at $10 and $50. StackBlitz's Bolt sells monthly token allowances, from 10 million tokens on Pro.
• You need a cost-plus overage. Cursor and Devin charge usage beyond the plan at "API list price." GitHub Copilot's AI Credits are consumed at each model's API rate.

When tokens are the wrong choice:

• Your buyers aren't technical. A bakery owner can't tell you how many tokens a week of social posts needs. Neither can most marketing managers.
• Your value is far above your cost. If a support reply costs you $0.007 and saves your customer a few dollars of staff time, a token price throws away most of the value you create.
• You'd be inviting a comparison with the vendor. Cost-plus turns you into a reseller. Your customer can look up the API price and ask why yours is higher.

Remember my telecom story. Phone companies didn't make customers think in electrical signals. They sold minutes, then plans, then "unlimited" with a fair-use clause. Twilio still prices its AI voice product per minute, like a phone call, because that's a unit buyers already budget for.

So translate. Charge per document, per conversation, per minute or per credit, and keep tokens as your internal cost unit. The credit-based pricing guide shows how to build that translation layer.

Are token prices going down or up?

Both, which is what makes this hard.

Going down. Epoch AI found the price of reaching a given level of performance fell about 40x per year, with a range of 9x to 900x depending on the task. Anthropic's flagship shows it: Opus 4 and 4.1 cost $15 in and $75 out per million. Opus 5.5 costs $4 and $20 (the Anthropic pricing teardown has the full history).

Going up. New frontier tiers sit above the old flagships at $10 and $50. Google is doubling two models' prices after launch. SpaceXAI raised its text-to-speech price from $4.20 to $15 per million characters in Q2 2026, according to AlloyPress.

Bills rising anyway. Agents and reasoning models use so many more tokens per task that total spend often climbs even as unit prices fall.

My advice? Price your product on customer value, and treat falling unit costs as margin you earn back over time. Never pass every price cut straight on to customers, and never assume a price you're paying today will hold.

Read the whole book free

The Subscription Playbook as a free ebook and audiobook. Australian business owners can claim a free hardcover, shipped free.

Claim the free hardcover Get the ebook

What to do next

1. Count your tokens. Log input, cached input and output tokens for your five most common tasks.
2. Cost each task with the formula above, at today's price and at any announced future price.
3. Find your expensive tasks. Agents, long documents and reasoning modes usually dominate.
4. Pull the levers: route simple tasks to small models, cache repeated instructions and batch anything that can wait.
5. Decide what your customer sees. Tokens for developers, credits or a natural unit for everyone else.
6. Set your price with the AI pricing method, then check your margin with the AI Gross Margin Calculator.

And if you take one thing from my telecom days: the customer who runs up a huge bill this month is not your revenue forecast. Build your plans so a quiet month still pays the bills.

Frequently asked questions

Get both books free

The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.

Leave this empty

First name

Email

CountryAustraliaNew ZealandUnited StatesUnited KingdomCanadaOther

What best describes you?Business owner (services)Business owner (retail / e-commerce)SaaS or AI founderHealth or allied health practiceTrades or home servicesAgency or consultantAccountant or adviserEmployee / marketerStudent / other

Yes, I'd like a free 30-minute Subscription Profit Review with a BPAG Accountants adviser (Australian businesses).

Your details go to Subscription Playbook and our sponsor BPAG Accountants Pty Ltd, who may contact you about the book and, if you ticked the box, your review. Unsubscribe any time. Privacy policy.

Send me the free books

Robert Coorey, MBA

Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.