What Is a Token? And Why Is It Suddenly on Every SaaS CFO's Radar?

What Is a Token? And Why Is It Suddenly on Every SaaS CFO's Radar?

By Alan Cox, CEO, Beakpoint Insights

Ask a ten-year-old what a token is and she will tell you it is the thing you use to play arcade games. Ask someone waiting for the subway and he will pull a small metal disc out of his pocket. Ask a poker player and she will point to the chips stacked in front of her. Ask a lawyer and he will tell you a token is a gesture, a token payment, a token of good faith.

The word has been carrying different meanings for different people for a very long time. It shows up in board games, transit systems, legal contracts, and religious ceremonies. It always means roughly the same thing: a small unit that represents something larger and has value in a specific system.

Which is exactly what it means in artificial intelligence. And understanding that simple idea turns out to be surprisingly important for anyone running a SaaS company in 2026

A New Kind of Token

When engineers building large language models needed a way to break human language into pieces a computer could process, they landed on the same ancient concept. A token in AI is a small unit of text that the model uses to read and generate language.

Not a word. Not a character. Something in between.

Think of it this way. The English language has roughly 170,000 words in common use. A computer that tried to work with whole words would need to memorize every one of them: every tense, every plural, every compound phrase, every proper noun, every word coined last Tuesday on the internet. The list would be impossibly large and would still fail the moment it encountered something new.

Individual letters, on the other hand, carry almost no meaning on their own. Processing text one character at a time would be like trying to understand a sentence by reading one letter every few seconds. The signal would be lost in the noise.

Tokens split the difference. They are the chunks of language that appear together most often in real text, the building blocks that carry just enough meaning to be useful without being so large that the system breaks down. Common words become single tokens. Longer or unusual words get split into familiar pieces. Punctuation, spaces, and numbers each have their own token rules.

The result is a working vocabulary of roughly 100,000 tokens, smaller than the full English dictionary but far more powerful as a processing unit.

It is worth noting that tokenization is not a new idea. Computer science has been breaking text into meaningful chunks since the 1950s, when compilers first needed to read programming languages. What is new is the scale and sophistication of how it is done in modern AI, and the fact that it now sits at the center of a significant and growing business cost.

Could You Put Tokens in a Dictionary?

This is where it gets interesting.

You could, technically, publish a list of every token in a model's vocabulary. It would look something like a dictionary, a long enumeration of entries, each one a fragment of language. You would find entries like "the," "run," "ing," "un," "2," "." "Beak" and "point."

But here is where the analogy breaks down. A dictionary gives you a word and then tells you what it means. Tokens do not work that way. Most tokens do not have a meaning on their own. They only mean something in combination with other tokens. "Beak" means something. "Point" means something. "Beakpoint" as a single concept only emerges when you see them together in context.

This is why tokens are less like a dictionary and more like an alphabet that has been turbocharged. The Roman alphabet has 26 letters. Those letters combine into words, and words carry meaning. The AI token vocabulary has 100,000 entries. Those entries combine into thoughts, and thoughts carry meaning. The scale is different, but the principle is the same.

If tokens are a language, it is one that was never spoken by humans. It was engineered, not evolved. Built by a machine studying billions of words of human text and finding the most efficient way to represent all of it. It is, in that sense, the most human language that no human has ever used.

And here is something that surprises most people. Each major AI model has its own tokenizer, its own distinct vocabulary of chunks. Claude and ChatGPT do not speak the same token language. The same sentence fed to each model may be split differently, producing a different token count and a different cost. For SaaS companies running multiple AI providers, this matters more than most people realize. You cannot directly compare token costs across models without accounting for the fact that they are counting in different units.

The Odd Things That Become Tokens

The rules of tokenization produce some genuinely surprising results.

Common emoji earn their own tokens. The crying-laughing face, the thumbs up, the heart. These appear so often in internet text that many models treat them as single indivisible units. Your users' AI feature is literally spending tokens on emoji.

Frequently appearing URLs and domain suffixes become tokens too. ".com" is almost certainly a single token in most models. A URL that appears millions of times in training data, a common API endpoint or a frequently cited website, may be encoded as one token even though it looks nothing like a word.

Programming punctuation sequences follow the same logic. Strings like "=>" or "::" or "{{" appear constantly in code and earn their place in the vocabulary. A code-focused AI feature burns tokens on syntax characters that carry no meaning to a human reader but are deeply familiar patterns to the model.

The underlying principle is always the same. It is not about looking like a word. It is about appearing together constantly in text. Frequency is the only criterion that matters.

Why Tokens Cost Money

Here is the part that brings this back to earth for a SaaS CFO.

Every time a user interacts with an AI feature in your product, types a prompt, pastes a document, asks a question, the AI model has to process every token in that input and generate every token in its response. Both directions cost money. The AI provider charges your company for each one.

Input tokens, what the user sends in, run at one rate. Output tokens, what the AI sends back, typically cost more because generating language is computationally harder than reading it.

The math sounds trivial at first. A fraction of a cent per thousand tokens. A single conversation might cost a tenth of a penny. Who cares?

The answer is: anyone running a SaaS product at scale, with hundreds or thousands of users interacting with AI features all day long.

Because tokens are not created equal across your customer base. A user who types a short question and gets a short answer burns a few hundred tokens. A user who pastes a fifty-page contract and asks for a detailed analysis might burn fifty thousand tokens in a single session. Both pay the same monthly subscription fee. Only one of them is running up your bill.

Token consumption is driven almost entirely by what each user chooses to do. It is behaviorally variable in a way that makes it uniquely difficult to forecast and uniquely important to measure.

The Token Is the Unit of AI Cost

The subway token and the arcade token share one important property: they make an invisible cost visible and countable. You cannot ride the subway for free. Every trip has a price, and the token is how you pay it.

AI tokens do the same thing for language processing. Every word typed, every response generated, every document analyzed, it all has a price, and the token is how it gets measured.

The difference is that in a SaaS company, the tokens are being spent by your customers on your platform and the bill lands on your desk. Until you can see which customers are spending how many tokens on which features, you are running a subway system where some riders take one stop and some ride all day, and you have no turnstiles.

That is the problem Beakpoint solves. Not just for AI costs, but for every layer of infrastructure spend. Cost per customer. Cost per feature. Cost per token. Visible, attributable, and actionable.

The token has always been a small thing that represents something larger. In AI, what it represents is your margin.

Beakpoint Insights connects infrastructure cost to business outcomes, giving SaaS companies cost-per-customer and cost-per-feature visibility built natively on OpenTelemetry. If AI costs are climbing and you cannot explain why, we would like to talk.