Someone posted a bill. The internet lost its mind.
Peter Steinberger, the Austrian developer behind OpenClaw who joined OpenAI in February, published a screenshot of his API usage dashboard showing $1,305,088.81 in OpenAI spending over 30 days — 603 billion tokens, 7.6 million requests, all driven by a team of three people and roughly 100 Codex instances running in the cloud. On a single day, May 15, his account logged nearly $20,000 in spend and 206,000 requests. The top model driving the bill: GPT-5.5.
OpenAI is footing the tab. Which is where this story gets interesting.
What the Agents Are Actually Doing
This isn’t vanity compute. Steinberger’s fleet isn’t running to generate marketing copy or fill a knowledge base. The hundred-odd Codex instances are doing real engineering work — reviewing pull requests, scanning commits for security vulnerabilities, deduplicating issues, writing fixes, and monitoring benchmarks for regressions. Some agents are even listening in on team meetings and opening PRs for features discussed in those calls.
That’s a genuine software development pipeline, automated end-to-end, operated by three people. Whether the output justifies the bill is a separate question — but the work itself reveals what Steinberger is actually testing.
His stated goal is deliberate: explore what software development looks like when token costs aren’t a constraint. Everything built in that process remains open source.
The Arithmetic Nobody Wants to Face
Steinberger clarified after posting the screenshot that the $1.3M figure reflects Codex’s “Fast Mode” pricing, which burns credits at a significantly higher rate than standard execution. Disable Fast Mode, and the raw API cost drops to around $300,000, which is still, by any reasonable standard, a staggering monthly bill for a three-person project.
That $300K figure is revealing differently. A single $200/month Codex Pro subscription provides roughly $5,000 to $6,000 in API-equivalent value per billing cycle, meaning OpenAI is subsidizing inference costs well below what the actual API would charge to drive developer adoption. Steinberger’s bill, paid by OpenAI, essentially makes the subsidy visible at scale.
The AI coding tools market is built on exactly this kind of pricing math. Codex, Claude Code, and Cursor are competing aggressively for developer mindshare, all three absorbing inference costs below their actual rates to lock in usage habits. OpenAI’s shift to token-based billing in April made those subsidies more transparent — but also more variable for anyone running at volume.
The question isn’t whether Steinberger’s team is doing impressive things with 100 agents. It’s whether this economic model survives long enough for those agents to become normal.
The Uncomfortable Proxy
Steinberger’s offhand comment after the bill circulated — “After turning off fast mode, my spending fell below the cost of one engineer” — sparked the real reaction online. Because that framing is exactly what makes investors and developers simultaneously excited and uneasy.
One comment from X, quoted widely in coverage of the story, cut to the chase: “Bro, you better show something that $1MM worth of engineers couldn’t do, or this might be the beginning of the advertising for the frontier lab bubble bursting.”
That’s not cynicism — it’s a legitimate demand. The AI industry has avoided the “bubble” label largely because the technology demonstrably works. Code gets written. PRs get reviewed. Security holes get flagged. But the economics of that work are still deeply artificial. Labs absorbing near-term losses to build usage patterns is a familiar playbook, and the playbook eventually requires revenue.
OpenClaw itself has had a turbulent public arc — from wiping a Meta AI director’s inbox to prompting Nvidia to develop a competitor. Steinberger has consistently framed the project as a stress test, not a product. That framing holds. But it only holds while someone else is paying.
What to Watch
The real signal in this story isn’t the bill. It’s what happens when the subsidy ends — when token costs reach a floor that actually reflects infrastructure, and developers have to make real decisions about how many agents they can afford to run.
Steinberger already knows the answer: with Fast Mode off, one engineer’s worth of spend, and a much smaller army. That transition, from 100 agents to the number a real budget allows, is where the AI development revolution will find its actual shape.
Related: OpenAI Is Copying Anthropic — And That’s the Real Signal of Who’s Winning AI
