Peter Steinberger, the OpenClaw creator who joined OpenAI in February, posted a screenshot Friday showing his team had spent $1,305,088.81 on Codex tokens over 30 days. The bill covered 603 billion tokens and 7.6 million requests. OpenAI, his employer, covers it.
What 100 agents actually do
Three people. About a hundred Codex instances running in the cloud, around the clock. They review every PR, scan commits for security holes, deduplicate GitHub issues once fixes ship, and ban spammers. Some monitor benchmarks and dump regressions into Discord. Others spin up temporary environments, log into services, and record video of the fixes they wrote.
A subset listens in on team meetings. When someone floats a feature in conversation, an agent opens a PR before the call ends. Whether that's useful or expensive theater depends on who's reviewing the PR.
The $300K asterisk
The headline number deserves a closer look. In a follow-up post, Steinberger noted that the $1.3 million figure reflects Codex's Fast Mode, which burns credits at a much higher rate. Disable Fast Mode and the raw API cost lands closer to $300,000. Still real money. Still roughly 60 times what a $200 Codex Pro subscription delivers in API-equivalent value, by Tom's Hardware math.
That gap is the actual story. OpenAI sells Codex Pro at $200 a month while the unsubsidized compute for a power user runs into six figures. Steinberger sits at the extreme end of that curve, but the distortion is structural. Every frontier lab is currently absorbing inference costs to lock in distribution.
"People freaking out over my AI spend. What nobody sees: part of what excites me so much about working on OpenClaw is that I'm trying to answer the question: how would we build software in the future if tokens don't matter?"
It's a tidy framing, and the question itself is reasonable. The catch is that tokens do matter, just not to him. Three-person teams running 100 agents in parallel without a sponsor don't exist yet, which is why the experiment is interesting and also why it's not a template.
What the dashboard doesn't show
The top model on the bill was GPT-5.5, dated April 23, 2026. On May 15, the day the screenshot went out, his account logged $19,985.84 across 206,000 requests. The screenshot itself came from CodexBar, Steinberger's own menu bar app, which tracks usage across 16 providers including OpenAI, Anthropic, Cursor, and Gemini.
What none of it captures: how much of the agent output shipped, how often the same bug got filed ten different ways by parallel scanners, or what the three human reviewers do with the firehose. Steinberger has said in past talks that he rejects "dark factory" auto-merging, which means somebody is still reading. At 206,000 requests in a day, that's its own form of work.
OpenClaw's next State of the Claw update is expected later this year. That's when the output numbers, if there are any worth publishing, will surface.
