Peter Steinberger, founder of OpenClaw, the AI assistant that became one of the fastest growing open-source projects in GitHub history, posted about a cleaner way to view API costs in CodexBar. What the internet latched onto was the number sitting inside the screenshot. A $1.3 million token bill in 30 days. Holy cow. That was my reaction when I first saw it too. But then I dug deeper and found his explanation of what was actually going on.
The screenshot that stopped the scroll
The dashboard showed $1,305,088.81 spent in 30 days, 603 billion tokens processed, 7.6 million API requests, and nearly $20,000 spent on the single day the screenshot was taken. The top model driving most of that spend was gpt-5.5.
Steinberger was not posting to flex. He was sharing an update to CodexBar, a macOS menu bar tool that tracks API costs across providers including OpenAI, Codex, Claude, Cursor, and others. The bill just happened to be visible inside the screenshot. The internet did the rest.
How the reaction played out
The replies came fast. Most people assumed this was one person’s individual API bill. Developer Dev Bredda pushed back in the thread pointing out that if this were just Steinberger’s personal spend, the full organisational bill covering everything he described in his workflow would be closer to $5 to $10 million. The implication being: something does not add up.
Steinberger’s reply was short. “That’s the org.” He later added that the org was roughly three people for most of the period the stats covered.
So the $1.3 million is real. The “one person burning cash” framing was not.
What were those tokens actually doing?
This is the part that got buried under the reaction. In a separate post, Steinberger broke down exactly what his team runs with those tokens. Around 100 Codex agents operate continuously in the cloud. Not chatbots. Agents that take actions.
Every pull request gets reviewed automatically. Every commit gets scanned for security vulnerabilities. Old GitHub issues get revisited, matched to recent fixes, and closed with references. Duplicate issues get identified and grouped into reports. Agents spin up temporary machines, recreate bug scenarios, record before and after videos, and attach them to pull requests. There are agents that read new issues and, if they fit the project’s documented vision, open a pull request automatically. Another agent then reviews that pull request. There are even agents that listen in on team meetings and start building features in real time while the discussion is still happening.
That last one is worth sitting with for a second. An agent attending your meeting and writing code before it ends.
What it actually means
A three-person team is running what would traditionally require a much larger engineering department, and the token bill is the infrastructure cost that makes it possible. Steinberger himself framed it as a deliberate experiment: how would you build software if token costs did not matter? The answer his team is living right now is faster, leaner, and with far less repetitive human work.
Token costs are falling. What costs $1.3 million today will cost significantly less in two years. The patterns being built at the frontier always find their way to the rest of us eventually.
FAQ
Was the $1.3 million really spent by one person? No. Steinberger clarified in the replies that the bill belongs to the organisation, which was roughly three people for most of the period shown in the dashboard.
What were the tokens actually being used for? Around 100 Codex agents running continuously, handling code review, security scanning, issue management, automated pull requests, performance monitoring, and even live meeting participation.
What is CodexBar? CodexBar is a macOS menu bar app that tracks API spending across multiple AI providers in one place. Steinberger posted the screenshot to show off a UI update to the tool, not to share his bill.
Leave a Reply