OpenAI’s Steinberger Puts Last Month’s Token Spend Near $10M

Peter Steinberger, the creator of the open-source AI agent OpenClaw and a member of technical staff at OpenAI, estimates spending around $10 million in tokens running AI workloads last month. The Register reported on 21 October 2026 that Steinberger said so on stage at Cloudflare Connect 2026 in San Francisco, in a conversation with Cloudflare’s chief technology officer, Dane Knecht.

Tokens are the units AI companies use to charge for the text a model reads and writes. The Register wrote that the money is “all coming out of investors’ pockets”, and Steinberger added a caveat: “What people don’t always understand is not all of that is useful.”

Business Pill · COST PER TASK

A short Business Engineer explainer on cost per task: what it costs to get a job finished, not the price of one word. It teaches the general idea only and says nothing about the figures in this piece.

What Was Said on Stage

• The budget. Knecht noted that Steinberger has acknowledged having “a virtually unlimited token budget, which not all of us in the room have,” according to The Register.
• The caveat. Responding to the audience’s surprise, Steinberger said: “What people don’t always understand is not all of that is useful.”
• The infrastructure. “I feel like the main challenge a lot of people still have is that our infrastructure structure hasn’t really caught up yet,” Steinberger said. “Like why is this still using my computer? Shouldn’t this all be in the cloud?”
• The future. Asked whether everyone would run hundreds of agents in a few years, Steinberger said: “I think everyone will have some sort of intelligence that runs and thinks and helps them throughout the day or even predicts what could help you in the future.”

Sources: The Register (Thomas Claburn), 21 October 2026; The Decoder, 16 May 2026; The Next Web, 23 October 2026. The May figure is a reported API bill; the October figure is a spoken estimate, so the two are not measured the same way.

The Earlier Figure From May

The Decoder reported on 16 May 2026 that, in 30 days, the OpenAI API bill hit $1.3 million for 603 billion tokens and 7.6 million requests, that the top model was GPT-5.5, and that “OpenAI picks up the tab.” Against that bill, an estimate of about $10 million for last month is roughly 7.7 times larger, by our arithmetic, though one is a reported bill and the other a spoken estimate.

Steinberger joined OpenAI in February, The Next Web reported, and has spoken about token use since. The Next Web quoted a June meetup in San Francisco: “Last year I was constrained by token. Now I fixed that by joining OpenAI.” At Y Combinator’s Startup School in August, asked when agents will run all the time on their own, Steinberger said: “That’s not so much a tech problem. It’s more a token problem,” according to The Next Web.

Agents Checking Agents

Knecht also asked how to decide what data to give an agent. Steinberger said part of the answer depends on the model, because so much of model behaviour is not governed by the harness, the code that wraps the model, The Register reported. “A lot of decisions that I built into OpenClaw early on now can change because the models are so much better,” Steinberger said.

Even with prompt injection risks reduced, Steinberger said, “you might not want to give your agent all the permissions.” The suggestion was a second agent that grants permissions: “You want to have another agent that can give it permissions or not and then this agent has to negotiate with another agent.”

On open source, Steinberger said pull requests should show how the code was reached. “If I see your thoughts, how you got from A to B, my trust in the output is way higher than if you just throw the finished work across the fence,” Steinberger said, according to The Register.

Why It Matters for Agent Budgets

The Next Web set the figure against Accenture, which it reported caps its staff’s Claude spending at $29 a month. About $10 million in a month works out to roughly $330,000 a day over a 30-day month, by our arithmetic. As we read it, Steinberger’s caveat that not all of it is useful points to the measure that matters when companies budget agents: what the spend finished, not how many tokens it used.

Connecting the Coverage

Earlier today, our Claude Max limits piece reported SemiAnalysis measurements showing one Max 20x account got a 20% smaller weekly budget than another on the same plan. And our harness scaling piece covered a paper in which training an agent across 18 harness setups raised a small open model’s average score from 29.81 to 39.04. What’s new: an OpenAI member of technical staff puts one month of AI workloads at about $10 million in tokens.

The Structural Read

As we read it, the figure measures access more than output: Knecht described “a virtually unlimited token budget, which not all of us in the room have,” and The Register wrote that the cost comes out of investors’ pockets.

As we read it, the gap between spend and useful work is the planning variable: Steinberger said not all of that spend is useful, which is the difference between tokens bought and tasks finished.

As we read it, the harness is moving into the model: Steinberger said many decisions built into OpenClaw early on can now change because the models are so much better, and suggested a second agent that grants or withholds permissions.

Three Implications

FOR AI LABS The Decoder reported in May that OpenAI picks up the tab for the API bill; as we read it, the lab is paying for heavy agent use by its own staff.

FOR COMPANIES BUDGETING AGENTS The Next Web reported that Accenture caps staff Claude spending at $29 a month; about $10 million a month is roughly $330,000 a day, by our arithmetic.

FOR AGENT BUILDERS Steinberger said “you might not want to give your agent all the permissions” and described an agent that grants permissions to another agent.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Subsidized AGI Economy.

The framework argues: “The negative margin is not a leak. It is a procurement budget. The labs are not pricing the product; they are buying an input that is not for sale anywhere else: agentic workload signal at the frontier of difficulty.”

As we read it, Steinberger’s month fits that frame: an OpenAI staff member running agent workloads at about $10 million of tokens, a cost The Register says investors carry. The framework was published in June 2026 and does not discuss Steinberger or this talk.

What Is Not Established

The $10 million is Steinberger’s spoken estimate as reported by The Register, not a published bill; the reports we read do not break it down by model, provider or task. The Register says the cost falls on investors and The Decoder said in May that OpenAI picks up the tab; the reports we read do not say how much of last month’s spend was billed at list prices.

We read The Register, The Next Web and The Decoder’s free opening and did not contact Steinberger, OpenAI or Cloudflare. Nothing here is investment advice.

The Bottom Line

OpenClaw’s creator, now at OpenAI, estimates spending about $10 million in tokens on AI workloads last month, against a $1.3 million OpenAI API bill reported in May. Steinberger also said not all of that spend is useful.