Skip to content

November 21, 2026 · Edition #108

Cheaper AI, bigger bills.

The price per token keeps falling, yet every AI plan is moving to weigh your requests. Vague ones weigh the most.


Every AI subscription is moving to the same rule: the heavier your request, the more it counts.

If you use Claude, you already know it. A long session, with big files and the most powerful model, eats your limit far faster than a quick question. Codex and ChatGPT's Work mode work the same way: the bigger and more complex the task, the faster you hit the ceiling.

In December, the rule reaches the place where most people use AI: the ChatGPT chat window.

Plus, Pro and Business will move to one allowance, shared with Codex and measured in tokens.

(A token is a small piece of a word, about three quarters of a word in English. Everything a model reads and everything it writes is counted in tokens. It's the unit AI is sold in.)

Quick chats with Instant, the fast model, won't count. The thinking models will, and the bigger the job, the more it weighs. The price of the plans stays the same.

If you live in Codex or Claude Code, none of this is news. You've probably been watching a usage bar for a while.

December is when it stops being a power-user problem. The rule that started with the coders now reaches everyone who uses ChatGPT to think something through.

The easiest way to picture it is a buffet.

In the early days, AI plans worked like an all-you-can-eat buffet. You paid a flat price, and the restaurant counted your plates. It never looked at what was on them.

A buffet works because the price is set for the average appetite. Most people take a reasonable plate, and they quietly pay for the few who pile it high.

Then agents arrived. And agents pile it high, all night.

So, one restaurant after another, the scales are going in.

That should surprise you, because AI has never been this cheap. When GPT-4 came out in March 2023, it cost $60 per million tokens of output. Today, much better models cost a small fraction of that, and the price keeps falling every year (Letter 103).

Every bite gets cheaper, and the scales keep spreading anyway.

Why? And what should you do about it?

That's what this letter is about. Once you see why the plates are being weighed, you'll also see why it's good news for one kind of person: the one who knows exactly what to order.

Why every restaurant is installing a scale

Start with the simplest math there is. A bill is a price multiplied by a quantity.

The price is falling fast. For the same level of capability, a token gets about ten times cheaper every year.

So if the total keeps going up anyway, only one thing can explain it: we're using far more tokens. Faster than the price can fall.

Round numbers make it obvious. At ten times cheaper a year, the price per token falls about 100 times in two years. Now say that in the same two years, you went from a few chats a day to agents running for hours, and your usage grew 300 times. Divide by 100, multiply by 300: your bill is three times bigger.

There are three reasons.

The work changed shape. A chat answers you once. An agent is an AI that works through a task on its own: it reads, plans, tries, fails, retries and checks, and every one of those steps burns tokens. Anthropic measured this on its own systems. An agent used about four times more tokens than a chat. A team of agents working together used about fifteen times more.

Models now think in tokens. Reasoning models write out a long chain of thinking before they answer you. You usually don't see it, but it's counted like any other text the model writes. The harder the question, the longer the thinking.

Cheaper things get used more. That's the Jevons effect from Letter 103. When something gets cheaper, we find a hundred new uses for it, and the total goes up.

Now do the buffet math. The restaurant priced the average plate. If a chat is one plate, an agent session is four, and a team of agents is fifteen. And it can keep going all night while you sleep.

The average appetite broke.

You could see it happen.

In June 2025, Cursor, the AI code editor, moved its $20 plan from a number of requests to a monthly credit charged at the real price of the tokens. A lot of developers found out when the extra charges showed up. The CEO apologized publicly and the company refunded the surprise bills.

A month later, Anthropic added weekly limits on top of its five-hour ones. The reason it gave: some people were running Claude Code, its coding agent, in the background, around the clock. Fewer than 5% of subscribers would be affected.

Those 5% were the guests who never left the table. And they were using AI exactly the way everyone is now telling us to use it: hand the work to an agent and let it run.

And the pressure is only going up. This week, Anthropic went public at a valuation of about $1.5 trillion, the largest IPO on record. A public company has to explain its margins to the market every quarter.

(To be clear, I'm not judging anyone here. I use Claude every day, and a lab that makes money is a lab that's still around next year.)

An open buffet is a great way to fill the room. Once you have to show a profit every quarter, the scale becomes the most important tool in the kitchen.

That's why I think the scale is here to stay, and why I expect it to get stricter.

When plates are weighed, the precise eat better

So far this sounds like bad news. For some people it is. Stay with me, because this is the part I most want you to take away.

Anthropic published something fascinating about its own research agent. They looked at why some runs did better than others, and found that the number of tokens spent explained about 80% of the difference.

In plain words: on hard problems, more tokens buy a better answer.

A token is a small piece of thinking. Give a model more room to think, and it explores more options, checks more, and catches more of its own mistakes.

Tokens are thinking you can buy.

So the real question a scale asks you is: where is thinking worth buying?

And the answer depends almost entirely on you.

Picture two people with the same plan, the same model and the same job: a proposal for a client.

The first one does what users learned to do at the buffet. They type "write a proposal for this client" and attach the whole shared drive. The agent reads everything, guesses what matters, and comes back with something polished and generic. Gorgeous garbage.

So they hit regenerate. Then "make it more convincing." Then they paste an old proposal and say "more like this." Six rounds later, they have something usable, and a big piece of their week's allowance is gone, because every round sent the agent back through the whole shared drive.

The second one spends ten minutes before ordering. Who the client is. The problem the client is actually trying to solve. The three points that will win them. What a good proposal looks like in this company. The two documents that matter, and only those two.

Then they hand it to the agent once.

The second person gets the better proposal, and the agent spends its tokens on the hard part: the argument.

When only the plates were counted, knowing what you wanted barely changed what you got out of your plan. Once they're weighed, it's the whole game.

That's what I mean when I say your expertise is GPS coordinates (Letters 70 and 99). A vague request sends the model wandering across everything it knows, and wandering is weighed by the token. A precise request gives it the shortest road to the part of its knowledge where the experts live.

It's also why I keep saying the prompt is never the bottleneck. You can polish the wording all day. If you don't know what you want, the model can't guess it for you.

I've always said AI amplifies the person operating it. It takes whatever you bring and makes it bigger, clarity and confusion alike.

The scale makes the amplification visible.

The habits we picked up at the buffet

Now the uncomfortable part, because it's about all of us.

We learned AI at the buffet, and we built habits that only make sense when nobody weighs the plate.

Regenerate until something looks right, without really reading the first answer. Paste the whole folder "just in case." Tell the agent to "figure it out" and walk away. Re-explain your business from scratch in every new conversation, because nothing is written down.

(Be honest: how many times did you hit regenerate this week?)

On a scale, each of those habits eats into your allowance. And when the allowance runs out, you wait for the reset, upgrade your plan, or buy credits.

In companies, it goes further, because there the plates turn into invoices. I think AI spending is about to follow the path of the cloud. When companies moved to the cloud, the bills surprised a lot of them, and it took years to learn which team was spending what, and why.

Soon someone in finance will look at the AI line and ask the only question that matters: what did we get for it?

The useful answer is the cost of a finished piece of work. Per proposal sent, per ticket resolved, per report someone actually used.

The teams with written processes and clear briefs will have a good answer. Their agents spend tokens on the hard problems and bring back better work.

The teams running on organizational folklore (Letter 78), the unwritten rules that live in four people's heads, will pay agents to rediscover how the company works. Every single time.

That's the cruel part of paying by weight. The messiest teams pay the most for the same help.

What I'd do on Monday

Look at where your usage goes. Claude already shows it under Settings, then Usage, and ChatGPT will show your combined allowance in the same place from December. Watch it for a week, the way you'd watch a new phone plan. You'll probably find that a few habits eat most of it: endless regenerations, giant pastes, agents left exploring with no clear goal.

Write the brief before you start the agent. The goal, the context, what good looks like, and what to ignore. Five minutes of your thinking saves a lot of the agent's. (That's the 5-Minute Rule: four questions on paper before you open any AI. It's the first layer of The Thinking Stack, my short course on the thinking that comes before every prompt.) Then keep that context in a file you reuse, so you never re-explain your business again (that's the folder from Letter 105, and what the AI OS Course helps you build). The lightest plate is the one you never had to refill.

Match the model to the job. Quick question? Use the fast model (on ChatGPT, Instant won't even count). Hard problem with real stakes? That's where a big thinking model earns its tokens, so give it room. In a company, the same logic works at scale: routine work on open-weight models (models whose files are published, so you can run them on your own machines, often for a fraction of the price), and frontier models for the hard part. That's how many of my clients divided their AI bills by ten (Letter 100).

Ask IT which models you're allowed to use. Before you move any work to a new model, check what your company has approved. IT and security teams have good reasons to be careful, and they'll probably be glad someone asked first.

Back to the buffet

In December, the biggest dining room in AI gets its scale.

Hardly anyone will notice. And that's how you know the buffet years are over: the generous period when a plate was a plate, whatever you piled on it, and nobody had to know what they wanted before sitting down.

The food will keep getting cheaper and better every year. What changes is who gets the best meal: the person who walks in knowing what they came for.

Knowing what you want is now the best deal on the menu.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)


↑ All editions Older →
Charafeddine Mouzouni — AI Scientist and Founder

Start with one email.