AI Tokens Explained: What They Cost, What They Measure, and Whether They Matter

Every AI tool is counting something and calling it a "token." You've probably run into the word before: on a pricing page or in an error message that cut off a long chat. Almost nothing explains what it's actually counting.

That matters because AI has a measurement problem. Its value is often indirect and difficult to compare with the way work happened before. Tokens offer something much simpler: a precise count of how much AI was used. But usage and value are not the same thing. And as more AI providers explore token-based pricing, confusing the two will become a very expensive mistake.

What is a token and why does it matter?

A token is the chunk of text a model reads or writes as one unit. It isn't a word, and it isn't a character. It's somewhere in between, and the exact boundaries depend on the tokenizer a given model uses. As a rough rule of thumb, OpenAI shared one token works out to about four characters of English text, or roughly three-quarters of a word. "Unbelievable" might get split into "un," "believ," and "able." Common words often get a single token to themselves; rare or made-up words get chopped into pieces (Nebius).

But the actual definition of a token matters less than the two key practical applications of them:

The first is usage limits. If you've ever hit a "you've reached your usage limit" message mid-conversation, that's about tokens, not messages or minutes. Providers cap you by counting input and output tokens. Input tokens are everything you send into the model: your prompt, any documents you've attached, the conversation history so far. Output tokens are whatever the model sends back, and the new prevalence of reasoning models, which use more tokens generating output, is partially why you may be experiencing these limits more frequently. 

The second is the context window, which is really just a model's memory. When companies launch a new model, they usually advertise its context window: the maximum number of tokens it can hold at once. That includes your prompts, conversation history, attached documents, and whatever it's generating in response. For perspective, Google’s 1 million-token models mean a context window of roughly 50,000 lines of code, or eight average-length novels, or transcripts of 200 podcast episodes, all held in the model's head at once. Its newest model, Gemini 3.5 Pro, launched with a window closer to 2 million tokens, reportedly the largest of any major model so far (Developers Digest). 

When volume becomes the goal

Most people only face tokens when they hit a usage limit. Some people have started chasing them on purpose. "Tokenmaxxing" refers to maximizing AI token usage and treating that volume as proof of productivity. It caught on because it's easy to measure and easy to gamify. In April 2026, an employee at Meta built an internal leaderboard called "Claudeonomics" that ranked coworkers by tokens processed; the highest-ranked individual averaged 281 billion tokens a month (The Information, via Inc.).

The scary part is that there’s a human cost to this kind of usage craze. Boston Consulting Group and UC Riverside researchers linked heavy, compulsory AI use to a form of cognitive overload they call "brain fry” associated with more errors and greater decision fatigue. That may help explain a broader tension we have written about before: AI adoption may be up, but enthusiasm is not. 

Tokenmaxxing can also carry a significant price tag. Uber reported it had burned through its entire 2026 AI budget in four months because of it. But the deeper problem is that token count is an input, not an outcome. It measures how much compute was consumed, not whether the result was useful, or worth what it cost.

The cost is real, even when you cannot see it

Most AI users are subscribed to a flat monthly or  yearly plan. These subscriptions separate everyday AI use from its underlying cost. Light users subsidize heavy ones, and you could run a model repeatedly without knowing whether a task consumed pennies or hundreds of dollars’ worth of compute. 

Research firm SemiAnalysis tested how large that subsidy actually is: they ran every paid tier from OpenAI and Anthropic through long, agent-style tasks until each hit its usage cap, then priced the same tokens at standard API rates. The gap was enormous. A maxed-out $200 ChatGPT Pro plan could represent something like $14,000 of compute, Claude's top tier around $8,000.

Whether those economics eventually push the industry toward metered pricing is still unclear. But the underlying tension is already here: token count does not tell you how much value AI created, while flat pricing makes it easy to forget that the usage still has a cost.

A token count is easy to quantify. Value is harder.

Most organizations still do not know which AI use cases justify their cost. Without a clear baseline for what the work would have cost without AI, token counts can become an appealing and precise stand-in, even when they say little about value.

Whether AI remains bundled into flat subscriptions or shifts toward token-based billing, organizations will need a clearer way to understand:

• How much AI their teams are actually using
• Which use cases are creating enough value to justify their cost
• How to monitor and control usage without discouraging useful experimentation
• Who should have access to which models, and under what governance
• When a more expensive model produces meaningfully better results
• When alternatives like self-hosting make sense, and what tradeoffs they introduce

Almost no one has them fully worked out yet. That is increasingly the work we do with clients: helping teams define what value means, understand where AI spend is going, and make better decisions about the systems they build and buy.