The Absurd #2: The Meter Runs Backwards

Dear reader,

In the first entry, The Ledger That Doesn't Balance, I wrote that my effort and my output had stopped reconciling. This entry is about a second column that stopped making sense: the bill.

On 22 September 2026, Anthropic released a new AI model, Claude Opus 5.5. I am writing this on 5 October, thirteen days later. In those thirteen days, the meter that measures how much AI we use went down. And what we built went up.

The numbers

We build with Anthropic's Claude models, on a monthly plan called Claude Max, at 200 US dollars a month. Each plan comes with a weekly allowance of AI use, measured in tokens, the small pieces of text a model reads and writes.

Before Opus 5/Fable 5.1, we were using 150 to 175 percent of a plan's weekly allowance. That is more than one plan can hold, so we ran two.

Since Opus 5.5, we use 30 to 35 percent of one plan per week. So we went from two plans to one.

Anthropic's own description is that the model performs at the level of its Fable 5.1 model on most work, and costs 40 percent less to run than the Opus before it. In our own use, it is almost on par with Fable. That is the ridiculous part: the better tool is the cheaper one.

What we built in those thirteen days

This is where the auditor in me wants the supporting schedule, so here it is, checked against the apps as they run today. Where something is not working or not finished yet, there is a note at the end. Since 22 September, eight new apps went live, and four older ones were extended. Together, the main code of these apps took in more than 400 saved changes (commits) in that time.

New and live since 22 September:

• whitehatstoic.com: our home: we build software, host it, run the tests that protect it, and research AI safety. It also carries this blog.

• evolved/vlvd.net: the platform that runs our projects, and is open to clients whose projects can live there too.¹

• Learn AI Alignment: bite-sized lessons on AI alignment and safety theory, from basics to advanced research, with saved progress and leaderboards.²

• vlvd.net Electrified: write notes and articles, follow the people you like, and talk about what they write.³

• Soft-launched - Wavecrate: royalty-free music and sound effects for videos, podcasts, games and ads, with a checkable license for every download.⁴

• Soft-launched - EnrichAI: five deep questions a day from Claude, and a truthful readout of your answers.⁵

• Soft-launched - Study Shot: snap a photo of a book page and get up to 20 multiple-choice questions, beginner to advanced.⁶

• Soft-launched - Cheapest Day Finder: find the cheapest days to fly, then book the flight, seats and bags in one booking, and manage it later - for now as a demo.⁷

Built before, and extended since:

• Cheap Recorder: press one button in Chrome and get your screen, your voice, your webcam and the other side of a call as one MP4, with a transcript.⁸

• Dear Superintelligence: an open collection of letters written by people to the advanced AI systems of the future, about what we value and why.⁹

• Soft-launched - Problem Graph: write your problems down, link them into a visual knowledge graph, and find solutions that worked for others.¹⁰

• IndividuationLab: research on AI alignment and human-AI coexistence through Jungian individuation.¹¹

Why it is absurd

Everything I learned about work says that more output costs more. More hours, more people, more machines, a bigger bill. When you build more, the meter runs forward.

Here it ran backwards. Eight new apps went live in thirteen days, and the bill for the intelligence behind them was cut in half.

The speed of this is, in a word, unfathomable. And we are not reading about it. We are inside it, every day.

What it gives, and what it does not

What it gives is plain. A small team can now run a portfolio of sites that once needed a company. An idea can go from a note to a live site in days; several on the list above did. And the cost of trying something is close enough to zero that the question changes from "can we afford to build this?" to "should this exist?"

What it does not give is just as plain, and an auditor has to say it.

Every site that goes live has to stay alive. Every app on the list above is one more thing that can break at any hour. That is why one of the eight new ones, evolved, is the platform our projects run on, and a large part of its job is keeping them up. The meter for AI went down; the meter for attention did not.

Tokens are not value. A lower bill tells you what the work cost, not whether it was worth doing. Eight new apps is a count, not a verdict. Some of them will matter, and some of them will not, and no model will tell me which.

Checking still takes a person. The model writes faster than any of us can read. Somebody still has to look at the page in a real browser, test it like an attacker would, and decide it is right before a stranger uses it. That part did not get 40 percent cheaper.

None of it is in our control. The model, its price and the weekly allowance belong to someone else. Next month they could all change, in either direction. The Stoic split helps here. How fast the tools improve is not in my control. What I choose to build with them, how carefully I check it, and whether I keep it safe, are.

Questions I am living with

• If the meter keeps running backwards, what is the real price of building something?

• When building costs almost nothing, does the scarce thing become maintenance, attention, or trust?

• What happens to the people and firms whose work was sold by the hour, when an hour holds this much?

• How many live things can one person honestly look after?

• And what will the next update do to everything we worked out in these thirteen days?

Until the next absurdity,

Miguel