Token rich, token poor

A few years ago the AI industry was sorted into the GPU rich and the GPU poor. The same sorting is now happening to engineers. Some of us work with what feels like an unlimited supply of tokens. Most work on a fixed plan and stop when it runs out.

From the outside the gap between those two groups looks like a gap in skill. It is a gap in money, and it compounds.

Token disparity is economic disparity, not skill disparity.

I am writing this from the rich side. My team at Durable spends way above average for a small engineering team in Canada, and more than many startups in North America.

This post is about engineers. The same applies across every function: product, design and product marketing.

Two engineers

Take two engineers with the same talent and the same care for the craft.

The first has tokens to burn. The $200 plan is only where their month starts. When it runs out, usage-based billing takes over, and the bill lands in the thousands. Before any code is written, a model argues through three architectures with them. They plan in one session and build in another. Every change runs the tests. A second agent reviews the work of the first, and a third drives the app in a browser and files what it finds. When a chore shows up for the fourth time they build a tool for it, and the tool spends tokens too. They never look at a meter.

The second pays for their own plan: $20, or $50, or $100 in a good month. Whatever the number, it is a ceiling with nothing behind it, so they ration. The good model is saved for the hard problem. The small stuff gets no planning pass, and no second agent reviews the first. Nothing runs overnight. When the limit hits on Wednesday, they wait for the reset or write it by hand. They care just as much. They cannot afford to be thorough.

Both of them ship code a model wrote. Generating code is the cheap part, and even a $20 plan produces plenty of it. So you would expect the slop to come from the first engineer, the one drowning in generated code. It comes from the second.

Slop is not generated code. It is generated code that nobody checked. Before these tools, writing code and understanding it were close to the same act: you could not type a function without thinking it through. A model breaks that link. It hands over code that looks right and may not be, and something has to do the checking that the typing used to do. That is either hours of a person reading, or tokens spent running the tests, driving the app, comparing the result against the spec, fixing what failed and running it all again.

The second engineer is short of both. They are expected to move at the speed of the model, and the checking loop is the first thing a ration cuts. The first draft ships.

More AI does not make AI correct. A second model that reads the first one’s work can be wrong in the same way. What tokens buy is attempts against things that cannot be talked into agreeing: a test that passes or fails, a compiler, a running app. Every loop is another chance to catch what is wrong, the way a team with tests and code review catches more than a team with neither. Someone still has to decide what right means, and no amount of tokens does that for you.

Rigor is the part that costs tokens. Abundance buys verification. Scarcity buys a first draft.

What abundance teaches

The first engineer is not hypothetical. It is how my team works. We spend tokens to think through architecture, to plan, to write the code and to QA it. We spend more of them on tools, harnesses, scaffolds and agents that do work for us, all of which use tokens heavily themselves.

What that buys over time is not more code. It is well structured, rigorous systems that are engineered entirely with AI, with far less slop than you would guess. Nobody on the team was handed that ability. We learned it by spending.

The public examples say the same thing. Nicholas Carlini at Anthropic had 16 agents write a C compiler: 100,000 lines of Rust that can build the Linux kernel. He says most of his effort went into what surrounds the model: “the tests, the environment, the feedback”. He also warns that the verifier has to be “nearly perfect”, otherwise the model solves the wrong problem. A small team at OpenAI built a product of about a million lines with zero written by hand, and named the work harness engineering: design the environment, specify intent, build the feedback loops.

None of this is prompting. It is harnesses, verifiers, scaffolds and agents that check other agents. That is a skill of its own, you only learn it by doing it, and the practice is paid for in tokens. You cannot learn to run sixteen agents on a plan that runs out on Wednesday.

The people furthest ahead treat an unused token as wasted practice. Andrej Karpathy says he feels nervous when he has subscription left over.

The gap in numbers

When engineers pay for their own tools, the typical spend is about $20 a month or a free tier. In the Pragmatic Engineer survey of 900+ engineers, around 30% were hitting their usage limits, and most of the complaints came from the $20 plans.

Above that sit the $100 and $200 plans, and for most people who pay their own way the $200 plan is the top. On the rich side it is the floor. When a plan runs out, the same tools keep going on usage-based billing at API rates, and that is where the money goes. One company in the Pragmatic Engineer’s latest round allows $2,000 per developer a month “through a Claude or Codex Max subscription, plus credits used”. Another caps it at $4,000. Scale’s survey of 38 companies found a median of $577 and an average of $1,040, and one seed-stage founder went from about $200 to about $3,000 per developer per month in six months.

So one engineer has $20 to $100 a month and a hard stop. The other has a few thousand and no stop. That is a gap of 30 to 150 times. Engineers have always had unequal tools, but a faster laptop was worth perhaps twice as much. This gap is of a different order.

An old pattern

Software is late to this. Other fields have always charged for practice.

An airline pilot is made of flight hours, and flight hours are bought. A first officer in the United States needs 1,500 of them, and getting there can cost well over $100,000. Military pilots need half as many, and the state pays.

A racing driver is made of seat time. Lewis Hamilton, a seven-time champion, says Formula 1 has become a “billionaire boys’ club”, and that starting again from a working-class family “it would be impossible for me to be here today”.

A record is made of studio time. The Beatles recorded most of their first album in a single day, about twelve hours in the studio. Four years later the same band spent months in the same studio on Sgt. Pepper.

The sociologist Robert Merton named the pattern in 1968: the Matthew effect, where those who already have are given more and the advantage accumulates. In each of these fields the thing that builds skill is metered, the meter runs on money, and afterwards everyone calls the result talent. Tokens are the newest meter.

Who is token rich

There is no single line between rich and poor. There are at least four, and none of them measures talent.

Where you live. The plans cost about the same everywhere. The salaries do not. In Stack Overflow’s 2025 survey the median back-end developer earns $175,000 in the United States and $22,086 in India. A $100 plan is under 1% of the first salary and over 5% of the second. The $200 plan is 11%, and usage-based billing on top of it is out of the question. Two engineers of equal ability, and one is token rich because of the economy they are paid in. Anthropic’s own data has usage per person rising 0.7% for every 1% of GDP per capita. India sits at about a quarter of what its population would predict, and Nigeria at a fifth. The report says it plainly: “the benefits of AI may concentrate in already-rich regions”.

Who you work for. Most tokens are bought by employers, not engineers. One company treats them as tooling and never mentions the bill. Another caps everyone at a $20 seat. The engineer chooses neither. It is decided by the company’s bank balance and by what its leadership believes, and a generous token budget is becoming a job perk. A freelancer, a student or someone between jobs pays for every token themselves.

Where you are in your career. Later in a career the money is there and the employer says yes without asking. Early on, the budget comes with a job, and the job is the thing that is getting harder to land. Stanford’s payroll data has employment for 22 to 25 year olds in highly AI-exposed occupations about 19% behind their peers in less-exposed ones. Experienced workers show no comparable gap.

Who you know. People with relationships at the foundation labs get early access and higher limits. Even the free programs follow reputation. When Anthropic gave away six months of Claude Max, it went to maintainers of projects with 5,000+ stars or a million NPM downloads, and OpenAI matched it the following week. Generous, and aimed at people who already made it.

These stack. An engineer at a funded startup in Vancouver, well into their career and with friends at the labs, is rich on all four. An equally good engineer freelancing from Lagos is poor on all four. Nothing about ability separates them, and every month the first one gets more practice than the second can buy.

Spending is not the skill

The obvious objection is that burning tokens proves nothing, and it is correct. Companies built leaderboards that rank engineers by tokens consumed, and people gamed them. Jellyfish, which tracks engineering teams, found that the top 10% of engineers are about twice as productive as the median while spending ten times the tokens.

So the marginal token is not worth much, and there is no magic number a team should hit. But look at who gets to find that out. You learn where the waste is by spending first. The token rich get to overshoot, see what was wasteful and settle on what works. The token poor never see the curve at all.

There is a second objection: that leaning on AI makes you worse. Anthropic ran a trial with 52 engineers learning a new library. The ones with AI scored 50% on the quiz afterwards, against 67% for the ones who coded by hand. But the pattern inside the AI group matters more. The engineers who handed the work over, the writing or the debugging, averaged under 40%. The ones who used it to understand, with follow-up questions, explanations and conceptual questions, averaged 65% or more.

The pattern that teaches you something is the one that costs more tokens. Asking why, and then asking again, is a luxury on a meter. Rationing pushes you towards “just give me the code”, which is the pattern that leaves you worse off.

What I take from it

If you run a team: tokens are tooling, not a perk and not a reward for seniority. Give everyone the same open budget. It is the cheapest training you will ever buy.

If you are token poor: ask about the token budget before you take a job, the way you would ask about salary. Until then, spend what you have on understanding and verification, not on more code. Ask why. Make it write the test.

If you are token rich: say so. It is easy to look at an engineer on a capped plan and see someone who has not kept up. Most of the time you are looking at someone who was never given the reps.

A skill gap closes with practice. This one widens with practice, because the practice is the thing for sale.

Thank you Logan for feedback on drafts of this post.