What Is a Token Worth? · Kerrberry Systems

TL;DR

A token is my lumber now. If I’m building with AI, tokens are the raw material, and a business is supposed to know the cost of its raw materials.
I can’t forecast a token the way a carpenter forecasts wood. Its value is hidden, quoted in percentages instead of dollars, and set by the people who mint it.
The sticker is flat. The worth might not be. The published rate never moves, but what the same job actually costs me in tokens, speed, and throttling might.
The strangest gap is the subscription. I pay a flat fee every month and I genuinely can’t tell you what I get for it, and I think that line is moving.
They already pull this lever in the open. When they advertise “2x” as a perk, that’s proof the dial exists. Assuming it only ever turns in my favor seems naive.
So I’m measuring it. A fixed task, my own plan, logged for weeks. Data in a post or two.

I was on the phone with a friend yesterday, another founder who leans on Claude Code the way I do, and he asked me something that sounded dumb. “Is a token always worth the same amount?” I laughed and told him of course, the rate’s right there on the pricing page.

Then I actually tried to answer him, and I couldn’t.

I know what a token costs on paper. I have no idea what one is worth to me, in my own account, at 9am on a Tuesday versus 11pm on a Sunday. Those turned out to be two different questions, and until yesterday I’d never once separated them.

Tokens Are My Lumber.

Think about a carpenter bidding a job. He knows what a board foot of lumber costs this week, and he can make a decent guess at what it’ll cost next month too. Lumber prices move, but they move in the open, on a public market, for reasons he can look up. He can price the job with real confidence, because the cost of his raw material is knowable.

Tokens are my lumber now. Every feature I ship, every bug I chase, every proposal I draft, I’m spending tokens to get there. They’re the raw material of this business the same way boards and screws are his. And a business is supposed to know the cost of its raw materials well enough to bid a job and still make money on it.

Here’s where it breaks down. I cannot forecast a token the way he forecasts a board foot of pine. Not because the math is hard, but because I genuinely don’t know what a token is worth from one week to the next, or even one hour to the next. And nobody’s told me why.

A Currency With No Exchange Rate.

Open the usage page in Claude Code and there’s no dollar sign anywhere. Just bars: a session bar, a weekly bar, a bar per model, filling as I work and resetting on a clock. It reads like a currency, except it’s a currency with no exchange rate. I can watch the balance move, but I can never see the price.

What a token is worth to Anthropic isn’t fixed either. I’d guess it moves with their own costs: electricity, cooling, however much of the data center is already spoken for that hour. What a token is worth to me, the guy paying a flat monthly fee and trying to bid a client job, is never quoted anywhere. I get a percentage, not a rate.

I got curious enough about this that I half-seriously looked into minting a token-pegged coin. Not a real currency, and not because I think a token has some intrinsic value you could actually trade on an exchange. I wanted to force a visible price onto something that’s priced for me but never priced to me. If the only way I can find an exchange rate for my own raw material is to invent one, the ambiguity is real, not imagined.

What I Think Is Going On.

Some of this I can say as fact, because Anthropic said it themselves.

This spring, Anthropic ran a real peak-hours penalty: on weekday mornings, 5 to 11am PT, your five-hour session allowance burned faster than it did the rest of the day, so you hit the wall sooner. It ran roughly late March to May 6, 2026, applied across Free, Pro, and Max, and by their own account touched something like 7 percent of users. The weekly cap never moved, only the session did. Then in May they rolled the penalty back, bundled with a doubling of the session limits.

Last month, Claude Cowork ran a “2x” promo, doubling the five-hour session limit for thirty days, June 5 to July 5, 2026. Worth noticing what didn’t move that time either. The weekly cap stayed exactly where it was. The perk doubled inside the day and stayed flat across the week, which is a specific shape for a promo to take if the point were simply “use more, on us.”

Zoom out and it’s a pattern, not a one-off. Weekly caps showed up in August 2025, the peak penalty landed seven months later, and it came off again in May alongside that limit increase. The Cowork promo followed in June. Four documented moves in under a year, each one announced, each one real.

I said in Stop Burning Tokens that the clock was a dead end, that timing your work to the hour didn’t pay off. I was talking about the sticker, the published per-token rate, and I still believe that part: it’s flat, and nothing since has changed it. This is a different question. The sticker not moving doesn’t mean nothing else does.

Past that, I’m guessing, and I want to be honest about which parts are guesses. My hunches: the first thread I open in a session is probably the cheapest one, with some kind of premium on every parallel thread after it, and the same task might draw less from my window at 2am than at 2pm, even though the sticker says otherwise. Limits themselves are never stated as numbers anymore, only as a bar and a percentage. Anthropic used to publish rough prompts-per-session figures and quietly stopped, so even the estimates floating around now are just that: estimates.

Here’s what convinced me this was worth measuring instead of wondering about. They’ve already turned a dial like this against peak-hour users once, and turned a different one in Cowork’s favor once, both in the open, both on the record. If a company is willing to move a lever one direction as a penalty and the other as a perk, I don’t see why I’d assume it only ever drifts my way when nobody’s watching.

So I’m Going to Measure It.

So I’m going to find out, on my own account, with real work instead of a lab test.

The plan is simple: one fixed task, small and repeatable, the same prompts against the same frozen slice of a codebase, every time. I’ll run it over and over, at different hours, on different days, sometimes alone and sometimes alongside other threads, for weeks. The task stays constant on purpose. If I let the work itself vary, I can never tell whether the platform moved or I just asked for more.

Everything else, I’m going to log and let run wild. That means tokens in and out, tokens per second, and which model actually answered. It means every throttle warning, every “approaching your limit” notice, every time I get slowed down or turned away. And it means the usage bars themselves, before and after each run, because those percentages are the closest thing to a price I’ve got.

I’m not trying to prove a conspiracy. I’m trying to answer the question my friend actually asked: whether a token is always worth the same amount. If it is, I’ll have a boring, honest null result, and I’ll say so. If it isn’t, I’ll have a chart.

Moving the Line.

I don’t know the answer yet. That’s the whole point of writing this down before I have a single data point.

Here’s what I do know. I run a business on tokens the way a contractor runs one on lumber, except I can’t tell you what a token costs me this week, and I genuinely can’t tell you what my own subscription buys me from one month to the next. That’s not a complaint. It’s just true, and saying it plainly feels like the finding, even before I’ve logged a single run.

Give me a few weeks. I’ll have real numbers instead of hunches, and I’ll write up whatever they say, including if the honest answer turns out to be that nothing’s moving, and I was wrong. Data in a post or two.

David Kerr is the founder of Kerrberry Systems. He builds custom software for businesses that want a straight answer about what their tools actually cost, not just what the sticker says. Find him on LinkedIn or GitHub.