TL;DR
OpenClaw creator Peter Steinberger spent $1.3 million in OpenAI API tokens in 30 days running 100 Codex instances on his open-source project. The bill, covered by OpenAI where Steinberger now works, represents 603 billion tokens across 7.6 million requests and provides the most concrete public data point on the cost of autonomous AI coding at scale.
Peter Steinberger, the creator of OpenClaw and an engineer at OpenAI, racked up $1.3 million in API costs in a single month by running approximately 100 Codex instances simultaneously on his open-source project. The bill, which covered 603 billion tokens across 7.6 million requests over 30 days, is the most visible demonstration yet of what happens when AI-powered software development is run without budget constraints, and of how quickly costs escalate when autonomous agents operate continuously at scale.
Steinberger posted a screenshot of the bill on X, showing $1,305,088.81 charged to the OpenAI API, with GPT-5.5 as the primary model. OpenAI is covering the cost: Steinberger joined the company in February 2026, and the spending is treated as a research investment in understanding what software development looks like when token economics are not a limiting factor.
Peter Steinberger X Post – source: X
What the agents actually do
The 100 Codex instances are not simply generating code. Steinberger’s three-person team has built an autonomous development pipeline in which AI agents perform a range of tasks that would ordinarily require a much larger engineering organisation. The agents review pull requests, scan commits for security vulnerabilities, deduplicate GitHub issues, write fixes, and open new pull requests based on the project’s broader roadmap. Others monitor performance benchmarks and flag regressions to the team’s Discord server. Some agents, according to The Decoder, even attend meetings and generate pull requests for features that come up in conversation.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The team also uses Clawpatch.ai, Vercel’s Deepsec, and Codex Security for additional bug and security analysis. The result is a development operation in which three humans oversee a fleet of AI agents that collectively perform the work of what would traditionally be a mid-sized engineering team.
The cost question
Steinberger has been transparent about the economics. He clarified that the $1.3 million figure reflects Codex’s “Fast Mode” pricing, which consumes credits at a significantly higher rate than standard execution. Disabling Fast Mode alone would reduce the raw API cost to approximately $300,000 per month, a 70 per cent reduction. At standard pricing, the operation would still cost $3.6 million a year, but the gap between the headline figure and the underlying economics illustrates how pricing tiers and execution modes can dramatically inflate reported costs.
When asked about return on investment, Steinberger said everything his team builds is open source and works with leading proprietary models as well as open-weight alternatives. “I’d say pretty high,” he said.
The figure is useful precisely because vendor marketing around AI coding tools rarely discloses raw spend and token volumes at this scale. Most enterprise teams planning agentic development tooling are working from projections and estimates. Steinberger’s bill is a concrete, public data point: 100 agents running continuously for 30 days on a large open-source codebase costs between $300,000 and $1.3 million per month depending on execution speed, before any optimisation.
Who is Peter Steinberger
Steinberger is not a newcomer to building developer tools at scale. The Austrian engineer founded PSPDFKit in 2011, bootstrapping a PDF rendering and annotation framework that became the standard for mobile document handling. By 2021, apps built on PSPDFKit were running on more than one billion devices worldwide, and the company raised $116 million from Insight Partners, its first outside investment after a decade of profitable, self-funded growth.
After leaving PSPDFKit, Steinberger began experimenting with AI agents as a personal project. OpenClaw, a self-hosted autonomous AI assistant that runs entirely on users’ own hardware, became the fastest-growing open-source project in GitHub history, crossing 302,000 stars by April 2026, overtaking React, Vue.js, and TensorFlow in a fraction of the time those projects took to reach similar milestones. The framework connects to tools people already use, including email, calendars, browsers, and messaging platforms from Slack and Discord to WhatsApp and iMessage, and allows AI agents to execute shell commands, manage files, and automate web tasks locally.
When Steinberger joined OpenAI, he announced that OpenClaw would move to an independent foundation to preserve its open-source character. “I want to change the world, not build a large company,” he wrote. “Teaming up with OpenAI is the fastest way to bring this to everyone.”
What it reveals about AI coding economics
The $1.3 million bill arrives at a moment when the economics of AI-powered development are a central preoccupation of the software industry. OpenAI recently opened ChatGPT subscriptions to OpenClaw’s 3.2 million users, allowing them to 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, concluding that the compute demands of autonomous agents running thousands of API calls per day were economically unsustainable under flat-rate subscription pricing.
The divergence between those two approaches reflects an unresolved tension in AI pricing. Subscription models are designed for human-speed interaction: a person typing queries into a chat window generates a predictable, manageable volume of API calls. An autonomous agent fleet generates orders of magnitude more, and the gap between subscription pricing and actual compute costs is the subsidy that either the provider absorbs or the user pays.
Steinberger’s bill makes that gap visible. At $1.3 million for 100 agents, the per-agent cost is roughly $13,000 per month, far more than any subscription plan covers. Even at the optimised $300,000, each agent costs approximately $3,000 per month. For enterprise teams evaluating whether to deploy agentic coding tools at scale, these numbers provide a baseline that no vendor’s marketing page will offer.
The broader pattern
OpenClaw’s trajectory, from a personal experiment to the most-starred project on GitHub to an OpenAI-sponsored research platform, reflects a broader shift in how software is being built. AI coding agents from DeepMind, OpenAI, and Anthropic are moving from proof-of-concept demonstrations to production deployment, and the question is no longer whether AI will write significant amounts of code but how much it will cost and who will pay for it.
The rise of AI-assisted development, from individual coding copilots to fully autonomous agent fleets, is compressing the timeline between a three-person team’s ambition and a large engineering organisation’s output. Steinberger’s setup, three humans and 100 agents, is an extreme version of what many companies will attempt at smaller scales over the next year.
The $1.3 million bill is not a cautionary tale. It is a receipt from the future, showing what it costs when AI development tools are used at full capacity, without the budget constraints that currently limit most teams to a fraction of what the technology can do. Whether that future is affordable depends on how quickly model inference costs decline, how efficiently agent orchestration frameworks manage token usage, and whether the security and quality challenges of AI-generated code can be managed at the speed these agents produce it.
