Tokens, Attention, and the New Labor Market

“Don’t you think maybe they are the same thing? Love and attention?”
— Sister Sarah-Joan, Lady Bird

How much do you value your time? We seem, as a society, to both value it very highly—paying 20% more to have food delivered than deal with the indignities of dining out—yet, at the same time, we throw it away for hours doing the very things we hate ourselves for but continue to do anyway.

Rather than thinking of the value of our time, maybe it's better to think about the value of our attention. No need to harp on what has become of our attention span as we trudge ever deeper into the big mucky of modern media. We already know the truth: we can’t pay attention for shit anymore. But as terrible as our attention is, companies are still able to squeeze value out of it.

Whether we make minimum wage or $250,000 a year, we’re paid to get shit done. Getting that shit sufficiently done requires our attention—sometimes given freely in pockets of deep focus, sometimes taken forcefully in mandatory, seemingly pointless meetings—and, more and more, that attention is demanded not by a manager or a routine set of tasks, but by an algorithm.

But how much is our attention really worth?

Today, our attention is valued based on time. But someone’s time isn’t really what we’re paying for—especially in salaried, white-collar work—it’s the work they can perform: the emails they can send, the designs they can generate, the critical strategic thinking they can bring to the 9th Zoom call of the day. Time is simply the closest, undifferentiated approximation of what’s actually being valued: output. Given how varied that output is from person to person and day to day, it makes sense to fall back on time as the thing we price. It also makes disparate forms of labor easier to compare.

Is there another undifferentiated unit of work that we can use to price our attention—one that’s more precise than time but still just as generalizable? With the rise of generative AI models in the last few years, the answer may be yes: tokens.

Tokens

A token is the basic unit of work that generative models like ChatGPT or Deepseek operate on. They charge based on how many tokens are used, whether you ask them to write a legal contract, generate software, or find a way to keep cheese from sliding off a slice of pizza. Here’s how much these models cost:

• Deepseek R1: $0.55 per 1,000,000 input tokens; $2.21 per 1,000,000 output tokens
• OpenAI o1: $15.00 per 1,000,000 input tokens; $60.00 per 1,000,000 output tokens

For comparison, a person can attend to roughly 6,000 tokens per hour and generate 1,500–4,500 tokens in that same time, depending on whether they’re writing or speaking (see references at the end for how this is estimated).

• At a salary of $200,000, the cost of our attention is $1,000 per 6,000 tokens.
• At the federal minimum wage, it’s $7.25 for the same number of tokens.

When we think about our attention and work in terms of how much information we can process, rather than the time spent, there’s a gulf the size of [insert tech CEO’s ego here] between what a real human charges and what a state-of-the-art AI model costs. Today, there’s also a similarly sized gulf between what a human can do in 1 million tokens and what OpenAI can—but that gap shrinks with each new model.

Tokens aren’t Art

“But wait!” you say. “This is stupid. Reducing all of a person’s work to tokens misses too much nuance. People aren’t computers. We’re not black boxes that take commands and spit out tokens. Our intelligence—our ability to reason about abstract and qualitative concepts—can’t be priced in tokens. We think about stuff while we sleep. We have aha moments in the shower. We create things that never existed before. We can make art.”

Sure we can. And sometimes we do. Even more rarely, we make a career out of it. There are artists. There are entrepreneurs who, through sheer force of will, bring new things into the world. But they’re the exceptions. We have an entire society built on the fact that this isn’t most people. Most people are trained—through our public education system—to fit into a mold, to apply for the job that’s posted, and to do the work that’s assigned. In its purest form, to tap “accept” on the gig that is offered.

The disparity between how we work today and how generative AI models work (and are priced) can be hard to grasp. Thinking about work in terms of tokens is an interesting thought exercise, but it’s just that, right? We’re not going to get paid by the token rather than the hour…right? We won’t have salaries that enforce a minimum token output…right?

In the early days of the Industrial Revolution, jobs so intrinsic to one's identity that they became last names—Weaver, Webster, and Fuller—were automated by spinning jennies and power looms. Careers that once required decades of mastery and conferred social status got atomized into stepwise tasks, performed by the unskilled and priced by the hour. The market that dictates how we work can change. It’s not hard to see how it might change again because there are already places where our attention is bought and sold at auciton.

The Attention Market

Facebook, Google, Netflix, and every other advertising-based platform make money by selling our captured attention. This isn’t news; it’s how our economy works. That it feels like the water we swim in just underscores how pervasive a pay-for-attention economy is.

Netflix charges about $30 per thousand viewers for a 15-second ad. Using our same estimate of tokens per hour, that’s roughly $7.20 per 6,000 tokens (from one person). Google Search can charge even more by directly matching users’ attention with companies best suited to monetize it.

By capturing the value of our attention, Facebook and Google have become two of the top ten most valuable companies in the world—just by repackaging and selling it. They don’t direct our attention; they merely capture and resell it. Imagine how valuable a company would be if it could leverage our attention to perform tasks each of us is uniquely good at.

Ad markets match interest and intent with goods that can be consumed. We also have businesses that match experts to the problems they can solve best, like GLG. But there are limits to how well these can function. The context needed to provide informed answers—and the timelines involved—keeps these services out of reach for most but the biggest corporations.

What if, when you asked ChatGPT a question, it automatically verified its answer with a human? What if the model knew enough about its own limitations to know when to call in an expert? What if you, as an expert, got a Slack message from ChatGPT (not a coworker), asking you to gut-check its response? What if each of these tasks had a dollar value, baked into your subscription price?

We already have model architectures that act as mixtures of experts, using chains of reasoning to decide when to pull from the internet or other tools. It’s not that big a leap, technologically, to implement a system that can direct our attention at scale—certainly not compared to the innovations Deepseek announced in the last couple of weeks.

Pay Attention

It seems clear these changes are coming. They feel inevitable. But if we accept that, we can still make predictions and better prepare ourselves:

1. The price per token of our attention will be relentlessly driven down by competition from generative models.
2. There will be a market built on the ability to target and leverage the attention of experts who still outperform generative models, and the price of their attention will resist competitive pressure.

Should we be excited, or resist these forces like breakers and luddites did in the past? I’m not sure. I’m terrified for folks graduating college right now. They have so much debt, and they’ll be competing directly with these models.

I hope these tools make it easier for everyone to tap into humanity’s collective expertise—and AI’s—to tackle the problems they truly care about, not just become customers of solutions for problems the wealthy decide are worth solving. Hopefully we’ll see a thousand new marvels on the scale of Wikipedia, rather than just higher margins for the few companies that manage to survive. I guess we’ll see.

Notes

Calculating the current price of tokens for a human being

Many a cognitive psychologist has turned their attention to the question of how long we can pay attention, and across multiple studies and a broad array of cognitive tasks, the answer turns out, unsurprisingly, to be: "it depends." But, we'll just say that you can focus for 20 minutes without needed a break, say for 10 minutes to refresh and recharge. Seems reasonable, and honestly aspirational given the needling interruptions that we all face despite our best efforts to perform focused work.

So for a given hour, we get 40 minutes of focused work. Throw in an hour lunch break, and in a 40 hour work week, you get about 23 hours and 20 minutes of someone's attention.

The average person reads about 250 words per minute, and the average speaking rate is 150 words per minute * Let's just say 200 wpm, for how much information we can consume with our limited attention. The average touch typist types about 50 words per minute.

Following, Open AI's documentation stating that 1 token is roughly .75 words **, the average person can then manage (very roughly) 6,000 tokens an hour on the input side and 1,500- 4,500 tokens per hour on the output side

References

• Carver, R. P. (1992). Reading Rate: Theory, Research, and Practical Implications. Journal of Reading, 36(2), 84–95.
• Jacewicz, E., Fox, R. A., & O’Neill, C. (2009). Articulation rate across dialect, age, and gender. Language Variation and Change, 21(2), 233–256.
• OpenAI Documentation
• Merchant, B. (2023). Blood in the Machine: The Origins of the Rebellion Against Big Tech. Little, Brown and Company.