The Price of Thinking

How four decades of human knowledge became a line item measured in tokens, and why understanding the economics matters more than picking a side.

My son Connor built his own AI agent on top of OpenClaw, the open source framework. He uses it to build apps, create projects, experiment. The problem is that running it on frontier models gets expensive fast.

That forced him to learn the economics. He spent time inside OpenRouter figuring out which models could do the work he needed at a price he could sustain. Not out of curiosity. Out of necessity. He got good at it. He knows which models to use for which tasks, where the open weights alternatives match the frontier options, and where they fall short.

I use AI every day too, but with higher subscription limits I was not tracking costs at the task level the way he was. I was thinking about capability. He was thinking about price.

A few days ago we were sitting down talking about the cost of running his OpenClaw agent and I asked him something simple: what does it actually cost a business to have a person do that same task?

He said just look up the salary.

You can look up a salary. That is easy enough. But a salary is not the full cost. Once you add employer taxes, pension, benefits, office space, management time, recruitment, training, absence, the true loaded cost of a person doing a task is 30 to 40% higher than the number on their contract. And that total, the real cost of human intelligence applied to a specific task, is genuinely hard to find in one place.

Meanwhile, AI pricing is published to the penny.

We spent the rest of dinner trying to figure out whether you could actually put the two side by side. Compare what a human costs against what a model costs, task by task, with real data.

Connor said four words. “I can build that.”

The number nobody publishes

If you want to know what it costs to use Claude Sonnet 4.5, you can find the answer in seconds. Three dollars per million input tokens. Fifteen dollars per million output tokens. Published. Updated in real time. Verifiable by anyone.

If you want to know the true loaded cost of a UK solicitor reviewing a contract, you will find the salary easily enough. But the full picture is harder. A solicitor on a median salary of £52,300 costs the business closer to £75,000 once you add everything up. A US lawyer earning $176,000 costs closer to $250,000. Those totals do not appear in any standard dataset.

One side of the equation is completely transparent. The other requires assembly. That gap is the reason the conversation about AI and work has been stuck in opinion rather than evidence.

The forty year invoice

For four decades, billions of people translated their professional knowledge into digital form. Every template, CRM record, project plan, spreadsheet, and email draft. We called it work. It was also training data. Models now reproduce the patterns of professional knowledge work at speeds and costs that make the original human effort look almost quaint.

The cost of a task used to be straightforward: the cost of the person doing it. Now the cost is split. The model processes the task. The human verifies and improves the output. That split changes the economics of every knowledge-intensive business. And almost nobody is measuring it properly.

The numbers

This week on the All In podcast, Jason Calacanis asked the question that every business leader is going to face in the next twelve months.

“When do tokens outpace the salary of the employee? Because you’re about to hit it. I’m about to hit it.”

Chamath Palihapitiya described giving developers “token budgets” the way companies once gave them software licences. He said his team needed to be at least twice as productive to justify the combined cost of salary plus AI. Otherwise, he said, “I’ll run out of money.”

These are not analysts writing research notes. They are operators running real businesses with real burn rates. And the question they are asking is the one Connor and I asked over dinner: what does intelligence actually cost?

So we measured it.

humanortoken.com tracks AI model pricing against professional salaries across more than a hundred business tasks. UK salary data comes from the Office for National Statistics. US data from the Bureau of Labor Statistics. AI pricing comes from OpenRouter’s live API. We update it weekly. Here is what the data shows.

A UK solicitor reviewing a ten-page contract: loaded cost, £112. Claude Sonnet doing the same task: twelve pence. A ratio of 935 to one.

A US copywriter drafting a 1,500 word blog post: loaded cost, $146. The AI: three cents. A ratio of 4,865 to one.

An executive assistant spending one hour on scheduling and correspondence: loaded cost, £25. GPT-4o Mini: one tenth of a penny. A ratio of 25,000 to one.

Those are today’s prices. Calculated from today’s salaries and today’s token costs.

But the ratios are only half the picture. The other half is quality.

We benchmark every task for the quality of the AI output. Contract review scores 65% confidence. The AI catches most standard clauses but misses contextual nuance. You still need the solicitor to check it. Blog post writing scores 80%. Good enough for a first draft. Not good enough if you need a distinctive voice. Presentation design scores 55%. The structure is there. The beauty is not.

The ratio tells you the cost. The quality score tells you the risk. Both numbers matter. Publishing one without the other would be dishonest.

The other side of the cost curve

Calacanis reported his AI agents costing $300 per day. That is $109,500 per year. For one agent. Running frontier models on complex, multi-step reasoning tasks around the clock.

Connor knows this first hand. Running OpenClaw agents on frontier models is the adoption challenge the entire AI community is wrestling with right now. Token costs scale with capability. The expensive models are not always better, but when you need deep reasoning, the cheap ones fall short. Matching the right model to the right task is a genuine skill.

Deloitte published research this month calling AI the fastest growing expense in corporate technology budgets. Some firms report it consuming up to half their IT spend. This is not going away. As agents become more capable and run for longer on more complex tasks, the token bill will become as significant a line item as headcount. Possibly more so.

The cost curve is not a single line. It varies by task, by model, by quality requirement, by volume, and by whether a human is still in the loop. Understanding that surface is going to be one of the most important operational skills of the next few years. Not understanding models. Not understanding AI. Understanding the economics of intelligence. What it costs. What it is worth. Where the crossover sits for each task. And how fast it is moving.

What the spreadsheet will never show you

I have managed people for thirty years. Hired, trained, promoted, and occasionally let go of professionals across technology, media, sport, and consulting. And I can tell you that the most important thing a great employee delivers does not appear in any task description.

A solicitor reading the room in a negotiation and knowing when to push and when to stop talking. A designer who hears “make it pop” and understands that the real brief is “make me feel confident presenting this to my board.” A project manager who knows that when a stakeholder says “I am fine with either option,” she means the first one and needs you to confirm it without making her say so. An executive assistant who books the morning flight because she knows the CEO thinks more clearly before lunch.

A model can draft the email. It cannot know that this particular client needs to hear it from you, in person, on a Tuesday.

These are not tasks. They are judgements. Built from years of context, relationship, culture, and the kind of accumulated human knowledge that was never written down because it lived in the space between people. No dataset captured it. No model learned it.

The numbers show the cost of the task. They do not show the value of the person.

Any honest conversation about the economics of intelligence has to hold both of those truths at the same time. The task is measurably cheaper with AI. The person is immeasurably valuable beyond the task.

The companies that understand this will build something extraordinary. The ones that see only the ratio will strip out their people and wonder why the culture collapsed six months later. The ones that refuse to look at the numbers at all will be outrun by competitors who did.

Where this goes

The honest conclusion is not human or token. It is human and token, in proportions that vary by task, by quality requirement, by context, and by how much you can afford to get wrong.

Within eighteen months, the majority of routine tasks currently performed by mid-level knowledge workers will cost less in tokens than in salary. Not because the models will be dramatically better, though they will be. But because token prices are falling at roughly 50% per year while salaries are rising at 4 to 6%. Run the numbers on any task where AI scores above 70% today and you will see the crossover approaching. For some tasks it has already passed.

That does not mean those people become redundant. It means the shape of their work changes. The task becomes automated. The judgement, the context, the relationship, the ability to know what the model got wrong and why: that becomes the job. The human moves from executor to orchestrator. And the organisations that figure out how to measure and manage that blended cost will outperform those that treat it as a binary choice.

This is going to become one of the defining conversations of the next few years. Beyond the model improvements and the capability gains and the tool access, beyond the trust questions and the safety debates, the economics of AI, the actual cost of running intelligence at scale, will sit at the centre of every business decision about people, technology, and growth. It already does for Calacanis. It already does for Chamath. It will for everyone else soon enough.

Why we built it

The site is early. The taxonomy is growing. The benchmarks are being refined. Some of the numbers will change as we stress test them and as models improve. That is the point. This is a living dataset, not a fixed report.

We built it because this conversation keeps happening without evidence. In boardrooms, on podcasts, over dinner tables. Everyone has an opinion on what AI costs and what people are worth. Almost nobody has the numbers side by side.

But the numbers are only the starting point. What we are really trying to understand is the transition that is already underway. Tasks are moving to agents. The economics of that shift are becoming measurable. And as more tasks move, the role of the human changes. Not downward. Upward. Away from executing the task and toward managing the systems that execute it. Away from being the person who drafts the contract and toward being the person who knows which contracts matter, which clauses carry real risk, and when the agent got it wrong.

That is the bigger question humanortoken.com is built to explore. Not just what intelligence costs today, but how the relationship between human talent and AI capability is evolving. Where agents are already good enough. Where they fall short. What the economics look like when a person manages ten agents instead of doing ten tasks. And what new skills, judgements, and roles emerge as that transition accelerates.

The conversation is not human or token. It never was. It is about how both work together, what that costs, and what it makes possible.

We think that conversation deserves evidence. Now it has some.

Know what intelligence costs.

https://www.humanortoken.com/

---

If this resonated, subscribe. The next piece unpacks what happens when the cost curves cross: which tasks are already past the point of no return, which never will be, and what the transition looks like for the people in between.

Craig Hepburn is an AI strategist and builder, Perplexity Fellow working at the intersection of technology, psychology, and system design to shape how AI operates responsibly in the real world.