Peter Steinberger has put a number on what an AI agent habit can cost at the extreme end. The creator of the open-source agent OpenClaw, now a member of technical staff at OpenAI, told the Cloudflare Connect 2026 conference in San Francisco on Wednesday that he estimates he spent around $10 million in tokens running AI workloads last month alone.
The bill is not coming out of his own pocket — investors are paying, The Register reported from the session — but the figure still drew audible surprise from the audience. Cloudflare CTO Dane Knecht, interviewing Steinberger on stage, noted that Steinberger has previously acknowledged having “a virtually unlimited token budget, which not all of us in the room have.”
“Not All Of That Is Useful”
Steinberger’s own response to the reaction undercut the brag. “What people don’t always understand is not all of that is useful,” he said — an admission that a large share of the tokens his agents burn through produce nothing worth keeping. Agents work by looping: reading context, calling tools, checking results and trying again, and every loop is metered. Run enough agents, around the clock, and the meter spins at a rate that turns a tool budget into a headline number.
That is the arithmetic now confronting every company deploying agents rather than demoing them. Consumer agent products such as Yelp’s Check For Me phone agent are sold on convenience, but behind each one sits the same per-token meter Steinberger described — and few buyers have an unlimited budget to feed it.
“Shouldn’t This All Be In The Cloud?”
Steinberger argued the industry’s infrastructure has not caught up with how agents are actually used. “I feel like the main challenge a lot of people still have is that our infrastructure structure hasn’t really caught up yet,” he said. “Like why is this still using my computer? Shouldn’t this all be in the cloud?”
The Register noted the tension in that position: cloud-hosted agents are exactly the expensive kind, and moving workloads off the user’s machine raises the privacy and compliance questions that local execution avoids. Cost, in other words, cuts both ways — local runs are limited by the hardware on a desk, cloud runs are limited by the bill.
An Agent To Authorize Your Agent
The session also covered how Steinberger decides what his agents are allowed to touch. Part of the answer, he said, is how much he trusts the underlying model, because so much behaviour is governed by the model rather than the harness around it. “A lot of decisions that I built into OpenClaw early on now can change because the models are so much better,” he said.
Even so, he warned against handing an agent every permission: “You might not want to give your agent all the permissions.” His solution is a second agent acting as gatekeeper. “You want to have another agent that can give it permissions or not and then this agent has to negotiate with another agent,” he said. “Many times the agents are like, ‘oh yeah, actually maybe I shouldn’t do this,’ or ‘there’s a simpler way that keeps it within the guardrails.’ So those two systems combined greatly increase my trust that I just can let the agent rip.” The pattern — agents checking agents — is the same defensive logic driving the security tooling now being built around agent swarms, including the agent-threat research published this month.
Knecht also pressed him on open source, where maintainers are giving up on accepting pull requests amid a flood of AI-generated code. Steinberger said the fix is provenance: contributors should show their reasoning. “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,” he said.
Why It Matters
Steinberger’s $10 million month is an outlier by design — an effectively unlimited budget at a frontier lab — but it prices the thing every enterprise agent rollout is quietly negotiating with: inference cost per useful outcome. Agents do not consume tokens like a chat window; they consume them continuously, in loops, retries and self-checks, and Steinberger’s own verdict is that much of that spend is wasted motion. Until the useful fraction rises, or token prices fall far enough to make the waste irrelevant, agent economics will favour narrow, well-fenced tasks over the always-on assistants he predicts are coming — a personal intelligence that “runs and thinks and helps them throughout the day,” as he put it, answering before the question is even asked. His on-stage remarks are a reminder that the person who built one of the most popular open-source agents believes both halves of that future: the agents are coming, and the bill is the problem.
Background
OpenClaw is Steinberger’s open-source AI assistant project, built around agents that act on a user’s behalf across tools and services; he has since joined OpenAI’s technical staff while continuing to speak publicly about agent design. His Cloudflare Connect appearance with Knecht ranged across agent costs, infrastructure, permissions and open-source maintenance, and was reported by The Register on October 21.
Source: The Register
0 Comments