BYOK: What AI Agents Cost and Why Markups Break Trust

Your AI vendor makes more money when you waste more.

Read that again, because it is the whole chapter. Most AI software you can buy today is priced so that the company selling it profits in direct proportion to how much compute you burn. Not how much value you get. How much you spend. Once you see that, you cannot unsee it, and it changes which questions you ask before you wire an agent into your business.

I want to walk you through what these things actually cost to run, because the numbers are less scary and more interesting than the marketing on either side suggests. And then I want to hand you the two questions that cut through all of it.

What A Token Actually Is

Skip the computer-science lecture. Here is all you need.

An AI model reads and writes in chunks called tokens — roughly three-quarters of a word each. “Rosebud Woman” is a few tokens. This paragraph is maybe eighty. When you give a model a task, it reads your instructions (input tokens) and writes back a response (output tokens), and you pay a tiny amount for each. That is the meter. Everything an agent does — reading a listing, writing a scraper, drafting an email, checking a number — is just tokens in and tokens out, metered like electricity.

The price per token is small enough that a single task rarely costs what a cup of coffee does. What makes AI bills real is volume — how many tasks, how big each one, and how many run at once. A founder running one agent a few times a day is spending pennies. A founder running ten agents in parallel around the clock is running a small power plant. Same meter, wildly different bill. The lever that matters is not the rate. It is what you point at the expensive meter and what you point at the cheap one.

The Old Math

The old math on AI cost was a single flat rate: pick a model, pay its per-token price for everything, and hope the bill stays reasonable.

The trouble is that not all work needs the same brain. Some of what an agent does is genuinely hard — deciding how to restructure an ad account, reasoning about brand voice, holding the strategy for a whole department. And some of it is dead simple — fetch this page, write a throwaway script to pull some numbers, reformat a list. If you run all of it on a top-tier frontier model, you are paying strategist rates to have someone alphabetize a spreadsheet. It works. It is also a quiet, steady overpayment on the ninety percent of agent work that is grunt work.

That overpayment is invisible until you look, which is exactly the condition a token markup thrives in.

The New Math

Route the work to the right brain at the right cost.

This is the single most important cost decision in running agents, and it is boringly mechanical once you name it. Cheap, fast, small models do the grunt work. A small model — a Haiku-class model, in my stack — will write you a scraper for a fraction of a penny. It will fetch a page, clean a data feed, summarize yesterday’s orders, tidy a listing field. None of that needs a genius. It needs a competent, fast, nearly-free worker, and there are excellent ones.

The frontier models — the expensive, slow, brilliant ones — go in the seats where judgment lives. The architecture decisions. The strategy. What I think of as the C-suite seats: the ones reasoning about the whole company, not executing a single narrow task. You run far fewer of those calls, and you are happy to pay for them, because that is where the model’s intelligence actually earns its price.

Right agent, right seat, right cost. My business plan names this as a core principle, and it is not an optimization I bolted on later — it is load-bearing. It is the difference between an agent workforce that costs pennies a day and one that costs like a bad SaaS habit. And it is the first thing a token markup has an incentive to hide from you, because every task you route to the cheap model is revenue the marked-up meter doesn’t collect.

The Misalignment

Now the part that matters most, stated as plainly as I can and as fairly as I can.

A lot of AI tooling works like this: the vendor buys tokens from the model providers at wholesale, then bills you for the tokens you consume, at a markup. Convenient. You don’t manage accounts or keys; you just use the product and pay your bill. I understand the appeal, and I am not going to pretend the people who build it that way are villains. Plenty of good software is priced on usage, and it can be an honest deal.

But look at the incentive it creates. When your vendor’s revenue is a markup on your token consumption, they make more money every time you burn more tokens. Their interest and yours point in opposite directions. You want fewer tokens, cheaper models, tighter loops, less waste — that is literally the new math I just described. They make more when you do the reverse. Nobody has to be acting in bad faith for this to bend the product. It just quietly shapes every default toward more usage: the setting that runs the bigger model, the loop that reruns one more time, the feature that “works better” when it thinks harder. The efficiency you are trying to build is the exact thing that shrinks their revenue.

That is the case against token markups, and it is not about any one company. It is structural. Ask it of any vendor: does this company make more money when I use more compute? If yes, then their incentives and yours diverge exactly at the point where you are trying to run lean.

The alternative is bring-your-own-keys. You connect your own accounts — your Claude Max, your ChatGPT, your API keys, or a free model on your own hardware — and the harness uses them directly. It never brokers a marked-up token. You see the real bill from the real provider. The company selling you the harness charges for the harness: the orchestration, the seats, the command center, the thing that turns loose models into a coordinated workforce. It profits when you succeed, not when you spend. In my plan I wrote that token brokering is where trust in AI software breaks down, and that refusing it is a brand position as much as a pricing one. I meant it as a line I will not cross.

The Receipt That Keeps Me Honest

Here is the part the BYOK evangelists — including me, five paragraphs ago — tend to skip. Owning your own keys does not protect you from yourself.

One afternoon, a single parallel swarm on my own setup burned $150 of credits in eight minutes.

Eight minutes. No markup, no villain — my own keys, my own models, my own mistake. I had fanned a task out across parallel agents, the loop went sideways, and because nothing stopped it, it kept spending until I noticed. That is the flip side of the power. The same architecture that lets me fire ten requests at once and run a company’s worth of work in parallel can also empty a bank account before you refresh the tab.

So I have to say the hard thing to my own pitch: cheap-model routing and bring-your-own-keys are necessary and they are not sufficient. What actually keeps the bill sane is safety equipment, and I mean that word precisely — the way a table saw has a blade guard. Not a feature you toggle on if you feel like it. Standard, on by default, load-bearing.

Three pieces of it:

Spend caps as hard backstops. A ceiling per seat and per run, enforced mechanically, so a loop that goes wrong stops itself instead of stopping when you happen to look. The cap is not there for the normal day. It is there for the eight-minute afternoon.

Routing as the default, not the exception. When the cheap model does the grunt work automatically, the baseline cost of activity is already near zero, so even a runaway on the cheap tier is a smaller fire.

Usage visibility. A live view of what each seat is spending, so a runaway shows up as a number climbing in front of you — a red dot, a rising figure — instead of a surprise at the end of the billing cycle. You cannot cap what you cannot see, and you cannot trust a system whose cost is opaque, including your own.

Model routing, context guardrails, and usage visibility are not optional niceties. On a system that can spend $150 in eight minutes, they are the brakes. I would not run agents against a real business without all three, and neither should you.

The Subscription Arbitrage

One more piece of money math, because it is worth real dollars and almost nobody spells it out.

For heavy, interactive, human-in-the-loop work — the kind where you and an agent go back and forth for hours, or a seat chews through a big job you’re watching — a flat-rate subscription is often dramatically cheaper than paying per token through the API. The subscription is a fixed monthly cost; the metered API bill for that same volume of back-and-forth can run well past it. So the move is to point your heaviest interactive lanes at a subscription you already pay for — a Max plan, a ChatGPT plan — and let the flat rate cap your exposure, while reserving metered API keys for the programmatic seats that run on a schedule and can’t live inside a chat app.

This only works if you bring your own keys. If a vendor is reselling you tokens, you can’t run this arbitrage at all — you can’t point a workload at your own Max subscription, because it isn’t your subscription in the loop; it’s their marked-up meter. BYOK is what puts the arbitrage in your hands. It is not just a trust position. It is money.

So What

You do not need to become an AI cost engineer to run agents. You need to ask two questions of anything you plug into your business, and hold the line on the answers.

Who owns the keys? If the answer is “we do, and we bill you for what you use,” you are renting your own compute back at a markup, and you have given up the subscription arbitrage and the transparent bill along with it. If the answer is “you do — bring your own,” you keep control of the accounts, the rates, and the ability to route work where it is cheapest.

Who profits when your usage grows? If your vendor’s revenue is a markup on your tokens, their business gets better exactly when you get less efficient, and every default in the product will lean, quietly, toward more. If they charge for the harness and leave your token bill alone, they win when you win. That is the only version of this where the person building your tools is rowing in the same direction you are.

Route cheap work to cheap models. Put the expensive brains only in the seats that need them. Cap the runaways, because your own keys will not save you from an eight-minute mistake. And bring your own keys — not as a feature you shop for, but as a stance you take.

The barrier here is lower than the pricing pages make it look. The tokens are cheap. The routing is mechanical. The two questions are simple. What is expensive is not asking them.