No. 20 soapbox
The Moat, Priced in Rice
I spent the summer learning to waste a machine's time on purpose, and the bill for that education depends entirely on which side of a moat you happen to stand.
This site is built with an AI. I have already confessed how and counted what it costs the planet. This is the third accounting, in money.
The meter and the bill
The receipt is the one from the t-shirt essay: fifty-one days of building, from the first commit of the puzzle pipeline to the morning of July 28, six billion The unit the model reads and writes in. One token is roughly three quarters of an English word.. What changes is the currency. Last time I converted them to electricity. This time I read the price list.
The maker of my model publishes sticker prices: five dollars per million tokens the model reads, twenty-five per million it writes, and a tenth of the reading price for anything replayed from The model re-reads the whole conversation on every turn. Providers store the repeated part and charge a discount for serving it back.. Run my fifty-one days through that list, as it stands in August 2026, and the meter reads about five thousand dollars.
To be plain about which numbers are which: the token counts are measured, off my own logs, and the prices are published. The guess is the mix. My logs do not record which model tier served every session, so I priced everything at the flagship rate. Price it all at the mid tier instead and the meter reads three thousand; at the top tier, ten. The argument survives the whole range, and the dial below lets you check.
For the builders: the meter math, bucket by bucket
Of the 6.05 billion tokens, about 5.72 billion are cache re-reads: at $0.50 per million that is $2,860. About 306 million are fresh reads: at $5 per million, $1,530. About 24 million are tokens the model wrote: at $25 per million, $600. Total: $4,990. I ignored the small premium charged for writing to the cache, and I rounded to five thousand. The tier dials move the total between roughly $3,000 and $10,000.
- What the meter would have read
- $5,000
- 51 days of tokens at list price, August 2026.
- Times what I actually paid
- 29×
- The plan, a flat $100 a month, came to $170.
- The meter reading, in rice
- 12,500 kg
- At $400 a tonne, the early 2026 export benchmark.
- Years of one person's rice in Bangladesh
- 50
- At 251 kg a year, the heaviest rice diet in the world.
Here is the part the meter does not show. I do not pay the meter. I pay a flat hundred dollars a month, which prorates to about a hundred and seventy across my fifty-one days. That is a nearly thirtyfold gap between what my usage lists for and what I was charged, and that gap is what this essay is about.
The tuition is the waste
Nobody taught me to run one robot, let alone several at once. There was no course and no manual that fit my project. I had a flat rate and a summer of experiments that mostly failed.
The receipts are all over this repository. The blog's front page exists because I had thirteen designs built and kept one. Whole games were designed, generated, playtested for weeks, and then deleted. Almost none of it shipped, and all of it taught me something no tutorial could, because every lesson came from my own project and my own mistakes.
The flat rate is what makes that tuition payable. On a subscription, a failed experiment costs nothing at the margin, so I ran hundreds of them. On the meter, every deleted design would have been a line item, and I would have run fewer experiments and learned slower. How fast I learned was set by how much failure I could afford.
There is a second fee the receipt leaves off. A failed experiment costs nothing at the margin and an evening no matter what, and my summer of them was a summer of evenings I could afford to spend this way. The subscription is the smaller tuition. The larger one is disposable time, and no plan, flat or metered, discounts that.
Priced in rice
Last time I priced tokens in t-shirts, and the unit did quiet work on my behalf, because nobody needs a t-shirt. So this time, rice, the thing a large part of the planet actually budgets in. It is dull, storable, and priced every day on a world market, which is the point.
The benchmark is Thai white rice, five percent broken, the standard export grade: four hundred dollars a tonne in January and February 2026, which is forty cents a kilogram at the dock. Retail is higher everywhere rice is eaten, so these are the biggest honest numbers rather than the kindest. Halve them for a rough market stall price and the shape of the argument holds.
In that currency: my hundred-dollar month buys two hundred and fifty kilograms. A person in Bangladesh, the heaviest rice diet in the world, eats about two hundred and fifty-one kilograms a year. One month of my plan is one year of one person's staple food, and the meter reading for my seven weeks of practice is fifty such years. At the world average diet, closer to a hundred and fifty.
The far side of the moat
Spread the subscription across the days of a month and it comes to three dollars and twenty-nine cents a day. The World Bank's poverty line, redrawn in June 2025, is three dollars a day, and by its count about 838 million people live under it. My learning budget, the one I have been describing as a gym membership, is larger than the entire daily budget of most of a billion people.
Free tiers exist, and they are genuinely free, and they teach the one thing the moat does not care about: chatting. What compounds is different. It is running the machine at scale, delegating, checking, and throwing work away, and that is exactly the part bought with waste.
And the bar keeps rising. Two years ago, using this machine well meant typing a good question into a box. In June it meant briefing one robot; by late July, running several at once and judging which of their work to keep. Each time the intelligence rises, the skill it takes to collect on it rises too.
So the moat compounds. Whoever can afford waste, in dollars and in evenings, learns faster, gets more productive, and can afford more of it. If an employer pays the bill, the moat is invisible. From under the poverty line, it is the whole horizon. The intelligence is nearly free. The tuition is the moat.
The view from the cheap side
I crossed for gym-membership money and did not notice crossing, which I suspect is what a well-built moat feels like from the inside. Token prices have fallen since I started paying attention and I expect them to keep falling. That will not fill the moat in, because the frontier moves, and the moat re-forms at whatever the frontier costs.
The t-shirt essay has a test I still believe: a comparison that changes what you do is analysis, and one that changes only how you feel is anaesthesia. Here is what this one changes. I will stop calling this skill talent. It is mostly tuition, paid in rice I never had to count.
Sources
- My own session logs, 7 June to 28 July 2026 The 6.05 billion token count and its cache/read/write split, as first published in the t-shirt essay.
- Anthropic API pricing The $5 / $25 per million token list prices and the one-tenth cache-read rate, as published August 2026.
- My own billing, August 2026 The $100 flat monthly plan. The $170 figure is that rate prorated over 51 days.
- IMF global rice price (Thailand, 5% broken), via FRED The $400 per metric ton benchmark for January and February 2026.
- Helgi Library: rice consumption per capita, Bangladesh 251 kg of milled rice per person per year, 2023.
- World Population Review: rice consumption by country The world-average figure of roughly 80 kg per person per year.
- Our World in Data: the $3 a day poverty line The June 2025 revision of the World Bank line, and the 838 million people under it as of 2022.