Peter Steinberger did not set out to make a point about AI economics. He just wanted a cleaner way to look at his own API bill, so he posted a screenshot of his OpenAI dashboard on X. It showed $1.3 million spent over the previous 30 days, powering roughly 100 autonomous coding agents running around the clock against the OpenClaw codebase.
The number went viral for the obvious reason: few people outside a handful of frontier labs have put a real dollar figure on what it costs to run a fleet of AI agents doing continuous engineering work, not a demo. For any business now budgeting for agentic AI, the breakdown behind that headline number is more useful than the number itself.
What is actually new
Steinberger’s agents were not idle. Across the 30-day window they burned 603 billion tokens over 7.6 million requests, writing and reviewing pull requests, scanning commits for security bugs, deduplicating GitHub issues, and flagging benchmark regressions in the project’s Discord. A team of roughly three people, Steinberger included, supervises the fleet rather than operating it hands-on, task by task.
The latest CodexBar update renders API costs wayyyy nicer. https://t.co/lJ4dxNHwzG pic.twitter.com/fCkWutJGzT
— Peter Steinberger xf0x9fxa6x9e (@steipete) May 15, 2026
OpenAI covered the bill as a research and engineering investment rather than billing Steinberger directly, by his own account of the arrangement. That detail matters for reading the headline number correctly: $1.3 million is what the work cost to run at full, unthrottled speed, not what a typical buyer would actually pay for comparable output.
What this means for teams running AI agents
The per-agent math is the part worth keeping. At the configuration Steinberger ran, each agent cost roughly $13,000 a month. Switching off Codex’s “Fast Mode,” which trades token efficiency for raw speed, cuts that by about 70 percent, down to roughly $3,000 to $4,000 per agent at standard rates. For a business evaluating agentic AI the way we outlined in our buyer’s guide to agentic AI lock-in, that is the range to model: nearly a 4x difference between the fastest configuration and a cost-optimized one, doing comparable work.
It also sharpens a question we raised when covering agentic AI’s move out of the pilot phase: lock-in risk is not only about data portability, it is about whether a vendor’s default configuration is quietly billing for speed nobody asked for. Enterprises that treat settings like Fast Mode as a procurement lever, not just a technical default, will see very different bills than teams that leave the defaults on. That is the same discipline we pointed to in margin, not growth, is the real scoreboard in AI software: the unit economics of AI tooling are decided in configuration screens, not press releases.
What to watch
OpenClaw itself is a useful bellwether. What began as Steinberger’s personal project has grown into a 501(c)(3) nonprofit foundation with more than 180,000 GitHub stars, and has reportedly drawn acquisition interest from both OpenAI and Meta. As the framework moves from personal experiment toward the kind of scaled deployment we tracked in shadow AI inside the enterprise, its published cost data will be one of the few public reference points for what running agent fleets actually costs once the pilot phase ends.
AI and business tech news, verified by a physician who reads the filings. One email a week, no noise.
