If you’ve followed AI news lately, you’ve probably seen headlines about the newest models from OpenAI, Anthropic, and Google being more “token efficient” than the versions that came before them. OpenAI recently claimed its latest flagship model can produce the same coding results using roughly half the tokens compared to earlier versions. Anthropic and others have made similar claims about their own newer releases.

But what does that actually mean, and does it matter for your business?

In plain terms, it means the model can do the same job while generating fewer tokens along the way, which, since you’re billed by the token, can translate into a lower bill for the same output. It’s part of a broader shift in the industry: after a period some are calling “tokenmaxxing,” where AI usage and spend climbed without much scrutiny, more companies are now watching what they’re actually getting for their AI dollar. But “more efficient” doesn’t mean “free,” and it doesn’t mean every task suddenly costs less. It depends on the model, the task, and how your team or software is actually using it.

That’s worth understanding before you assume a new model release automatically shrinks your bill, and it starts with knowing what a token actually is.

So what actually is a token?

A token is the basic unit an AI model uses to process language. Roughly speaking, one token is about four characters of English text, or a bit less than one word. “Accounting” might be one token. “Fractional CFO advisory” is closer to four or five tokens once you account for punctuation and word breaks.

Every time you send a prompt to an AI model and get a response back, the platform counts tokens on both sides, what you typed in, and what it generated in reply. You’re billed on the combined total, not just your output. Worth noting: output tokens typically cost several times more than input tokens, since generating a response takes more computing power than reading one, which is part of why a model that’s more efficient at generating fewer output tokens can have an outsized impact on your bill.

Measuring and managing token usage

For most business owners, the practical question isn’t the linguistics, it’s cost control. A few things worth tracking:

Most platforms provide usage dashboards. If your team is using AI tools embedded in other software (a CRM with an AI assistant, for example), the token cost may be bundled into your subscription rather than billed separately, which makes it harder to see and easier to lose track of.

Tokens are not equal across platforms

This is where a lot of confusion sets in. A token from OpenAI, a token from Anthropic, and a token from Google are not standardized units the way a gallon or a kilowatt hour is. Each company tokenizes text slightly differently based on their own model architecture. The same sentence might break into a different number of tokens depending on which platform processes it.

That means comparing “cost per token” across providers without adjusting for tokenization differences can be misleading. The more useful comparison is cost per output, meaning what it actually costs you to get a report generated, an email drafted, or a dataset summarized, since that normalizes for the tokenization quirks underneath.

Converting dollars to tokens

Pricing is typically quoted per million tokens, and it varies widely by provider and by model tier within that provider. A lightweight, fast model might run a few cents per million tokens. A frontier, high capability model can run into the tens of dollars per million tokens. Providers also often price input tokens and output tokens differently, with output typically costing more than input.

There isn’t a single universal exchange rate, the way there is between currencies. If your team is trying to budget or forecast AI spend, the more reliable approach is tracking actual dollar cost against actual usage patterns over a few months, rather than trying to back into a fixed conversion rate up front.

Where this leaves you

If your company has meaningful AI spend, whether that’s a few hundred dollars a month or a growing line item, it’s worth understanding what’s actually driving the bill before it grows into a line item nobody can explain. Newer, more efficient models may help, but they’re not a substitute for actually knowing your usage.