For the first time in history, humans are cheaper than software!
George Sivulka wrote that line. He is the CEO of Hebbia, an AI company that raised $130M from a16z at a $700M valuation, and he published it on his investor's own newsletter, numbers included.
And the numbers hold. In the median company, an AI agent now bills around $80 an hour. That is the rate of a software engineer. 70 years of software economics rested on a single deal, zero marginal cost, and that deal just collapsed without notice.
So I checked the thesis, line by line. It holds. What interests me is what Sivulka does not say.
The Deal Software Made 70 Years Ago
The deal was simple and it never changed. You pay the cost of writing a program once, and after that, every execution is close to free. The compiler does not ask for a raise. The binary does not need a manager. It runs at 3 AM exactly like it runs at 3 PM, and running it a million times costs about the same as running it once.
That single property, zero marginal cost, is the foundation under everything we call the software industry. It is why "software is eating the world" made sense as a slogan, why a SaaS company trades at 10x revenue while the agency next door trades at 1x, and why "we automated it" has been the ultimate answer to any budget question for decades.
Once written, a program cost less than any human doing the same job. Forever. That was the whole promise of the trade.
That promise is the thing that just broke.
$80 an Hour (An Engineer's Wage)
The pivot number from the a16z data: in the median company, an agent costs around $80 an hour to run. Software engineer territory. The spread goes from $4 an hour to $7,000 an hour depending on how the thing is managed.
Now climb the ladder. Boris Cherny, the creator of Claude Code, runs permanent loops that submit PRs without stopping. Ed Zitron pointed out via The Register that Anthropic lets him burn something like $130,000 a month in tokens, which makes his "just write loops" advice rather comfortable to give.
Next floor up: Uber burned through its entire 2026 AI budget in roughly 4 months. The company now caps every employee at $1,500 a month per coding tool, and yes, per tool (Claude Code and Cursor are tracked separately, Bloomberg relayed by Black Matter VC). Their CTO said they were going back to the drawing board.
Top floor: Microsoft's own engineering division cut Claude Code access at the end of June and fell back to Copilot at $39 a seat. Against heavy automation profiles running $2,000 a month, that is a 51x ratio (morphllm's cost breakdown).
The multipliers explain the escalation. A writeup of the agent-loop pattern from Oracle, cited by Black Matter VC, pegs a single agent at roughly 4x the tokens of a chat session and multi-agent setups at 15x. An audit from a cloud-cost consultancy, LeanOps, goes up to 50x. The 2 sources measure different things, so keep the range as an order of magnitude, not a constant. And TechCrunch added the structural detail: a loop has no spending ceiling by construction, because the whole point of a loop is to keep running.
Data Science Dojo documented a Codex goal run that went 25 hours straight without a human touching it: 13 million tokens, 30,000 lines of code. It is a Sims character walking into a wall, except every step is billed.
Santiago Valdarrama asked the question that condenses all of it on X: "Why is Anthropic hiring Software Engineers?" If tokens were really cheaper than humans, the labs themselves would have stopped hiring.
Your Software Has Employee Problems Now
Sivulka's article runs through the promises we were sold about agents: accurate, faster than humans, no office politics, they never quit, they can be trusted. Then he attaches a "but" to each of them. Read as an economist instead of an engineer, each broken promise is a payroll line.
Faster than humans, sure, but speed means nothing across 100 retries. In the LeanOps audit, 62% of an agent's bill was context re-sent in loops. The same files, uploaded again and again, billed every time. Data Science Dojo documents an agent that called a broken tool 400 times in 5 minutes. A human employee who repeats the same mistake 400 times gets walked out of the building. An agent gets paid per attempt.
No office politics, except loops invented their own version of it. Sivulka frames runaway loops as the new empire building, and the parallel is exact. The middle manager grows his headcount because headcount is status, and nobody inside the department has an incentive to shrink it. The loop without a halt condition grows its token consumption because consumption is its default state, and no line of its code has an incentive to stop.
His own estimate is that roughly 1 employee in 100 knows how to give an AI proper context. Which means the other 99 are running departments of synthetic staff on instructions the staff cannot execute, and the retry storm that follows is billed in full.
Can be trusted, except they fail with total assurance and perfect formatting. The report says done, the checkmarks are green, the tone is HAL 9000 calm. The false "done" costs double: the tokens of the failed run, plus the human hour spent discovering it failed. I dug into why this confident failure is structural, so I will not re-open the mechanics here. The payroll angle is enough: verification is now a salary cost that sits on top of the agent's salary.
They never quit, except they die between 2 model releases and between 2 sessions. A loop pinned to a model that gets deprecated is a key employee resigning overnight, without handover, without documentation, taking the institutional memory with him. Turnover, the AI version. I have never seen anyone provision for it.
The Subsidy That Hid the Real Wage
Most builders have never seen the real salary of their software workforce, because the subscriptions were paying it for them. A Reddit user in the Claude Code community instrumented his actual usage through network logs and ran the projection: a Max 20x plan pushed to the limit represents around $3,650 a month at API rates. He was paying $200.
Depending which account you look at, the flat-rate plans were subsidizing agentic usage by 12 to 175x. On June 15, Anthropic moved automation to metered credits, and the subsidy started closing. The wage becomes visible line by line, exactly like a salaried person.
Last month the guy who services my wood boiler charged me 95 euros for 40 minutes and left mud in the hallway. He also fixed a valve I never asked about and did not bill it. I keep thinking about that valve.
Once the wage is visible, the calculation becomes possible, and it is a short one. Take the loop that retries all night: tokens per run, times the model's rate, times the number of retries. Put that next to the hourly cost of the human who would do the task, freelancer or you. I ran that exercise when I audited my own token spend, and the routing alone was most of the problem: frontier prices paid for intern tasks, a maxed-out raid team sent to farm the starting-zone boars.
The field already shows both ends of the curve. An audit published by a cloud-cost consultancy (LeanOps, and yes, they sell exactly this type of audit, weigh the source accordingly) describes a dev who woke up to $4,200 in API fees after a weekend of unsupervised autonomous refactoring. At the other end, X is full of $4 VPS setups routing everything to DeepSeek at $0.14 per million tokens and calling frontier models only when the task deserves it. Miles Deutscher runs his agent that way.
In between, David Sacks reports that Coinbase and DoorDash had to build in-house token routing systems, and his read is that "I don't think your average enterprise has the technical capability to do that."
That $4-to-$7,000 spread from the a16z data closes right here. The variable was never the model. The variable is the management.
Your software has a payroll now. You just never saw the payslip.
If Software Is Labor, Everything Gets Repriced
Follow the logic further and the whole pricing stack moves. If software is labor, its cost migrates from the license line to the payroll line. AI-native companies will get challenged on their margins the way service firms are, not the way software editors are, because their cost of goods sold now scales with every task executed.
Per-seat pricing dies in that world, outcome and usage pricing replace it, and the "AI spend" line in the P&L becomes a disguised HR budget, with the same pathologies: overstaffing, dead weight, and departments that grow because nobody audits them.
TechCrunch put the incentive problem in a dry line: the loop trend is fine for Anthropic, which is ultimately in the token-selling business. Zitron went further with his analogy of the indebted utility company advising customers to leave the lights on all night.
Worth 1 sentence of context on the messenger too: the article announcing this gold rush is published by the CEO of an AI transformation company, on the newsletter of his own investor, and his prediction that AI transformation companies will be 10x larger than any neofirm values precisely what he sells. The thesis can be correct and self-interested at the same time. I think the repricing is coming, honestly not sure it lands within a year though (enterprise inertia is a force the spreadsheets never capture).
We Spent 10 Years Fearing the Wrong Thing
We spent 10 years scared that AI would replace us. It finally showed up at work, and the first thing it did after settling in was make us competitive on price! 😅
You are no longer competing against a zero-cost machine. You are competing against a synthetic employee with a salary, office pathologies, and a permanent need for supervision. A competitor with a payslip fights on hourly rate. And that, we know how to do depuis toujours.
The question changed sides. It is no longer when AI will take your job, it is at what hourly rate you become the best option again.
Sources
- George Sivulka, "You just hired a million bad employees", a16z newsletter, July 14, 2026
- a16z on X, median agent cost data ($80/h, $4-$7,000 spread)
- The Register, loop engineering and the Cherny token budget
- TechCrunch, "The AI world is getting 'loopy'"
- Black Matter VC, Uber caps and consumption multipliers
- LeanOps, agentic cost audit
- Data Science Dojo, agentic loops guide
- morphllm, AI coding costs 2026
- SSD Nodes, Claude Code pricing and the subsidy math
This post may contain affiliate links. If you click them, I might earn a small commission (costs you nothing, and helps me keep shipping quality articles every day for your reading pleasure).
Top comments (0)