OpenClaw creator’s $1.3 million OpenAI bill shows agent costs

Peter Steinberger, the creator of OpenClaw and now an engineer at OpenAI, said he spent $1,305,088.81 on OpenAI API usage in a single month while running about 100 Codex instances on his open-source project. The figure, which he shared as a screenshot on X, covered 603 billion tokens and 7.6 million requests over 30 days, according to the post and reporting by The Decoder.

The spend is one of the clearest public data points yet on what autonomous AI coding can cost when it runs continuously at scale. Steinberger said OpenAI is covering the cost and that the spending is being treated as a research investment in understanding software development when token economics are not the main constraint.

The setup behind the bill is not a simple code generator running in the background. Steinberger said his three-person team uses AI agents to review pull requests, scan commits for security vulnerabilities, deduplicate GitHub issues, write fixes, and open new pull requests based on the project roadmap. Other agents monitor performance benchmarks and post regression alerts to the team’s Discord server.

According to The Decoder, some agents even attend meetings and then generate pull requests for features discussed there. The team also uses Clawpatch.ai, Vercel’s Deepsec, and Codex Security for additional bug and security analysis.

Steinberger has been direct about the pricing side of the experiment. He said the $1.3 million figure reflects Codex’s “Fast Mode,” which uses credits at a much higher rate than standard execution. He said turning off Fast Mode would cut the raw API cost to about $300,000 per month, a drop of roughly 70 percent.

At standard pricing, Steinberger said the operation would still cost $3.6 million per year. The gap between the headline bill and the lower estimate shows how much execution mode and pricing tier can change the economics of agentic development.

When asked about return on investment, Steinberger said, “I’d say pretty high,” adding that everything the team builds is open source and works with leading proprietary models as well as open-weight alternatives. The point, he said, is less about a single bill and more about exposing the real economics of running AI coding agents without tight budget limits.

Steinberger is no stranger to building developer tools at scale. He founded PSPDFKit in 2011 and turned it into a widely used PDF rendering and annotation framework. By 2021, apps built on PSPDFKit were running on more than one billion devices worldwide, and the company later raised $116 million from Insight Partners after a decade of bootstrapped growth.

After leaving PSPDFKit, Steinberger began experimenting with AI agents as a personal project. OpenClaw, which runs on users’ own hardware, became the fastest-growing open-source project in GitHub history, reaching 302,000 stars by April 2026, according to the article.

The bill also lands during a wider debate over how AI coding tools should be priced. The article says OpenAI recently opened ChatGPT subscriptions to OpenClaw’s 3.2 million users so they can run autonomous agents through the Codex endpoint for $23 per month. Anthropic, by contrast, blocked Claude Pro and Max subscribers from using OpenClaw and other third-party agent frameworks, saying the compute demand from autonomous agents was too high for flat-rate pricing.

That tension matters because autonomous agents do far more work than a person using chat-based software. A human typing a few prompts creates a modest load, while a fleet of agents can generate thousands of API calls a day. Steinberger’s bill puts a number on that difference: 100 agents running for 30 days cost between $300,000 and $1.3 million per month, depending on execution mode, before any further optimisation.