When AI Gets Expensive, Will Humans Get Cheap Again?
OpenClaw creator Peter Steinberger posted his OpenAI bill. US$1.3 million in API tokens over 30 days. 603 billion tokens for 100 Codex agents running autonomously for a three-person team.
This is the showcase scenario. If autonomous AI agents are going to replace human teams, this is what it looks like. An overhead of US$1.3 million for a small software development team working in an agentic environment.
The Two Ways We Work With AI
I believe most people are not Peter Steinberger and this is how we use AI every day. Two ways of using it. Both produce output. Only one of them is actually thinking with you.
The first way is the one most people know. Open a chat window, ask a question, take the answer. AI breaks the question into components, retrieves what it needs, reasons across the parts, generates a response. This is information processing done well. Decomposition. Retrieval. Synthesis. Output.
The second way is different. You are not asking for an answer. You are asking AI to think alongside you on a problem. You challenge its first response. You feed it constraints it could not have known. You tell it where its assumption breaks against an outcome it has never seen. The output gets sharper not because the model got smarter, but because you gave it the context and information it could not infer on its own.
This is where AI proves its worth. Not as a source of answers. As a reasoning engine that needs direction.
In my daily working routine, I do mostly the second way. US$25 and 23% of my Claude Code Opus session, gone in the first 30 minutes. Thinking with AI is not a US$20 monthly chatbot subscription.
The Economics of Giving AI The Wheel
If 30 minutes of thinking with AI costs me US$25, what does a full operation cost when AI runs it?
Steinberger’s bill made the news. The economics behind it did not. US$1.3 million a month works out to roughly US$43,000 per day. For a three-person team. The agents do not take breaks, do not negotiate salary, do not need health insurance. They also do not stop spending.
The math sounds wild until you remember what an agentic operation actually is. Not a one-off query. Each agent is running its own loop, on its own task branch, at its own pace. None of them stops on its own. The agent receives an input, calls an LLM to reason about it, exchanges data with an API, evaluates the response, decides the next action, calls the LLM again. Each turn costs tokens. Multi-step workflows multiply the cost. 100 agents running in parallel multiply it again.
This is the part the autonomy narrative never mentions. Agentic operation is not free software running on your laptop. It is a metered service consuming compute every second it is awake.
When AI Gets Expensive, Humans Get Cheap Again
The irony in Steinberger’s bill goes deeper than the headline number.
US$1.3 million a month is the price of three engineers running 100 agents. The same money hires 70 senior engineers working full time. At some point the cost curve crosses, and the cheaper labor becomes human again.
This is not a thought experiment. Eventually the same conversation will be picked up by the newspaper headlines. The math is simple. An agent that costs US$13,000 a month to run does work that a junior engineer could do for less. A fleet of 100 agents costs more than the engineering team it was supposed to replace.
The autonomy narrative assumed AI would get cheaper as it got better. So far we see the opposite is happening. Models are getting more capable and more expensive at the same time. Reasoning consumes more tokens. Agentic loops consume more reasoning. Each capability upgrade is also a cost upgrade.
Somewhere in this curve, the spreadsheet flips. The CFO who approved the agent fleet last year will approve the human team next year. Not because humans got better. Because AI got expensive enough that the comparison stopped being obvious.
The Agent Does Not Know When To Stop
The cost problem and the cognition problem are the same problem.
An agent runs the loop because the loop is what it was built to do. It does not stop to ask whether the task is worth the tokens. It does not pause to consider that the same fix has been attempted three times and maybe the specification is wrong. It does not look at the bill and decide that the operation is producing more cost than value.
A human would. A junior engineer who burned US$43,000 in a day would be in a meeting the next morning. They would justify the spend or learn not to repeat it. They would notice that productivity measured in commits is not the same as productivity measured in shipped value. They would know when to stop.
The agent does not have that move. It runs until it is told to stop, and it is rarely told to stop because the dashboard shows it working. Working and producing value are not the same thing. The agent cannot tell the difference.
This is not a budget problem that better pricing fixes. It is a judgment problem. And judgment is the thing AI does not have.
Until AI Knows How To Throw A Curve Ball
A curve ball is the unexpected move. The one that breaks the pattern because someone read the situation and decided the standard response was wrong. It does not come from the playbook. It comes from instinct shaped by experience the playbook never captured.
AI does not have that move. It cannot reach outside the mainstream of its training to make a decision that contradicts what it has been taught. Ask it for the standard answer and it will deliver one. Ask it for the answer that breaks the pattern because you sense the standard is wrong here, and you will get a polished version of the standard anyway.
The unexpected move lives outside the training data. It cannot be retrieved. It can only be lived.
Until AI knows how to throw a curve ball, the agent will run as long as the tokens hold. It will produce output as long as the API responds. It will not stop to ask whether the work matters because it does not know what mattering looks like.
But you do. That is the job. That has always been the job.
