The Inference Bill That Never Stops

00 / The short version

The short version

01 / The surprise bill

The worry about layoffs was wrong. The bill is the real story.

In 2026, the headline was that language models would swallow white collar jobs whole, and companies cut staff to get ahead of it. Then the people who started the panic changed their tune. OpenAI's Sam Altman now says he is delighted to be wrong: he expected far more entry level jobs gone by now than actually are. Anthropic's Dario Amodei, who had warned of a white collar bloodbath, traded it for a gentler story about work multiplying instead of vanishing. An outside group, the Yale Budget Lab, ran the numbers and found no real jobs dent through early 2026.

Here is the part most people missed. A hosted model bills you by the token. A token is just the unit the model providers charge by, think of it like a taxi meter. The meter runs the whole time the model is working: reading, thinking, and writing. You pay for every word in, and every word out.

• The speed is real. But you buy it one token at a time, and the meter never stops. Frontier models are not cheap, and when the system quietly retries after a failure, you pay for that too.
• The sums get huge, fast. Uber handed agentic coding tools to about 5,000 engineers and blew its entire annual budget for them in four months, with some engineers running up $500 to $2,000 a month each. Its fix was a hard $1,500 a month cap per person. (June 2026.)
• A layoff is a one time choice you can see on paper. An inference bill is different: it is a cost that comes back every month, and it hides inside the word productivity until it becomes a problem.

02 / Our own real numbers

Meet Neos: one autonomous agent wearing eight hats

Neos is a real autonomous agent we built and put to work inside a real business: a construction and property operation with crews in the field, properties to track, books to keep, and a brokerage account to watch. Neos lives on a harness called OpenClaw and talks to the operator over the same chat app the crew uses. It never sleeps, never takes lunch, never complains. One catch: for every single task, it rents one of the most capable language models money can rent, Anthropic's Claude Sonnet 4.6, and pays by the word. Every task, a fresh charge.

And every task covers more than you would guess, because Neos is not one worker. It is one orchestrator wearing eight specialist hats, with 65 different commands spread across them.

What does that look like on an ordinary Tuesday? Every hour on weekdays, four email watchers sweep the inbox: bookkeeping notifications get turned into cards on the project board, a vendor's purchase orders get flagged, a key property contact never goes unanswered, and price drops on watched listings raise a hand. It pings the foreman in the crew chat for a quick yes or no on a material list, reminds the operator about quarterly sales tax deadlines, nags us when the brokerage login needs its weekly re-authorization, and checks the pulse on its own live trading dashboard. Then, at the end of each day, it sends the operator a work report with per task hour estimates, and every Friday at 5PM, a weekly wrap with next week's schedule. None of that is a demo. That is just its Tuesday.

So we did the thing most people never do: we tracked the bill, to the penny, across a stretch of real days. A busy weekday runs about $85 to $141. Add it up and Neos is on pace for $2,300 to $3,000 a month, roughly $28,000 a year. One recent day, June 9, hit about $286 before the market even closed.

Here is the part that made us laugh, then wince. Neos carries a giant instruction book. Its core handbook, soul file, and memory notes run about 9,000 words. The eight specialists' operating manuals stack on another 17,000. Add the catalogs for every tool it can touch and the full briefing Neos re-reads at every wake up comes to roughly 130,000 words, and it pays for every one of them, from page one, even when nothing has changed. Now multiply: 32 jobs run on their own schedule, and Neos wakes up about 166 times every weekday. That is a lot of re-reading.

03 / The fix we built

A notebook, a student, and an agent named Mila

The fix is a framework we built called Recall, and the agent we raised on it, Mila. Strip away the jargon and it comes down to two old fashioned ideas: a careful notebook, and a student who studies it every night.

The notebook first. Inside Mila, every job gets filed: what was asked, what worked, what it cost. Before any new job, the system runs what we can only call a deja vu check (have I seen this before?), and when the answer is yes, the old answer is reused for free. When a better answer comes along, it layers on top of the old one without erasing it, the way scribes used to reuse parchment. That is literally what we named it: palimpsest. Anything fresh or time sensitive, a price, a deadline, a live order, is never reused. That always gets a real, current answer.

The student is the part we love. The notebook doubles as a lesson plan, and the student is a small language model, 1.5 billion parameters, walnut sized next to the rented frontier model, that lives on a Mac sitting on our desk. Overnight, the Mac re-reads the notebook and tutors the student on it: a few hundred focused passes, targeted tutoring rather than re-schooling. Nothing is uploaded. Nothing leaves the machine. The only line item is electricity, and once the student has learned a job, every word it writes afterward is free.

04 / The replacement

Neos to Mila: the replacement that runs eight companies

We did not want you to take our word for it, so we did not ask you to. We built Mila as Neos's replacement: same kind of work, same days, but with the notebook and the student where Neos had only the meter. Mila is built on Recall, and she is no longer an experiment. She is in production, running the back office for 8 companies at once.

Neos's side of the board is measured, not guessed: it comes from the agent's own token level usage ledger over June 3 to 9, priced at public API rates. Treat those figures as lower bounds, since some logged sessions are incomplete. Mila's full cost numbers, measured the same way across all 8 companies, land in the next report, CSR-003. What we will not do is fake a scoreboard.

05 / Getting to almost free

Each round drives the bill down a little more

Here is the part that compounds, and it is our favorite part. Every job Mila does is one more page in the notebook, and every overnight training run teaches the student a little more of the business; the loop simply re-runs as the notebook grows. So with each round, fewer jobs need the expensive rented model, and the bill drops, round after round, closer to the floor. We will be straight with you: we are not at zero, and we will not pretend we are. But the trend only points one way.

06 / Staying safe

A person always says OK on the things that matter

Saving money is great, but not at the cost of safety, and especially not with an agent that teaches itself. So we keep firm rules, and these are not aspirations on a slide. They run in Mila today.

07 / The takeaway

You pay once to teach it. You keep what it learned.

The market mistook a billing problem for a jobs problem. The honest version of our story is not we zeroed out the inference bill overnight. It is a system, a notebook that never forgets and a student that studies it every night, that drives the bill down round after round, with a real deployment as the proof. You pay the frontier model once to learn your work. Then the learning lands in a notebook you own, inside a student you own, on a machine you own. Neos taught us the cost; Mila is how we take it back.

The goal is simple: own the thing that does the work, and take back the bill.

Neos's side of the board is on the page. Mila's full numbers are next. Watch this space: CSR-003 is where her measured ledger lands.

Sources / The outside numbers

Don't take our word for the market, either

Our own costs are measured in house, to the penny. The claims about the wider market are public reporting; here is exactly where to check them.

Every report starts with a real, measured baseline. The next one keeps score.