$1.3 Million a Month to Run 100 AI Coding Agents: Who Is Actually Paying?
$1.3 Million a Month to Run 100 AI Coding Agents: Who Is Actually Paying?
Peter Steinberger posted the bill, and then walked it back. On May 16th, the OpenClaw founder and OpenAI employee shared a screenshot from CodexBar, his menu bar API usage tracker, showing $1,305,088.81 in OpenAI API charges over 30 days. The breakdown: 603 billion tokens, 7.6 million requests, roughly 100 concurrent Codex instances. On a single day, May 15th, the tab hit $19,985.84.
The internet did the math. At that burn rate, three people were running more inference than most startups use in a year. The response on X was swift: _that is the annual salary of a senior engineer in San Francisco, spent not on salaries but on tokens_.
Steinberger's answer was precise. "I can disable fast mode and it's 70% cheaper," he replied. "So it's more like one employee." He frames the spend as research: an experiment in what software development looks like when token costs are removed from the equation. OpenAI, his employer, covers the tab.
The number is real. The question is what it means.
The Person Behind the Cost
Steinberger is not a typical indie developer burning through savings on a hobby project. He created OpenClaw, the open-source agent framework that sits at the center of a small but devoted ecosystem of developers who use it to orchestrate multi-agent systems. He is also, as of earlier this year, an OpenAI employee working on agent infrastructure.
That dual role matters. When a founder who also works at the lab running the most expensive agent experiment in public, sharing the receipts, the story is partly about technology and partly about incentives. Steinberger has every reason to demonstrate that heavy AI agent usage is defensible, even productive. His employer has every reason to underwrite that demonstration.
This is not a criticism. Labs subsidizing their own developer ecosystem is not new. But it means the $1.3 million figure is not a market price. It is an internal cost center, categorized as research. The actual market — enterprises, indie developers, startups — does not have OpenAI covering the invoice.
What 100 Agents Actually Do
According to The Decoder's coverage, Steinberger's three-person team runs the fleet across a set of well-defined tasks: reviewing pull requests, scanning commits for security vulnerabilities, deduplicating issues, filing bug reports, monitoring benchmark regressions and posting results to Discord. Some agents open pull requests based on the project's stated direction. Others listen to internal meetings and draft feature proposals.
The team also uses third-party services: Clawpatch.ai, Vercel's Deepsec, and Codex Security for supplementary bug and security analysis. The 100 concurrent Codex instances are the core infrastructure; the external tools layer on top.
The top model driving usage: GPT-5.5, specifically the April 23, 2026 snapshot. That is the most expensive frontier model in OpenAI's lineup. Running 100 instances of it, at scale, explains the token count.
The Fast Mode Variable
Steinberger's 70% cost reduction point is worth dwelling on. Codex Fast Mode is OpenAI's higher-throughput inference tier for coding tasks. The trade-off is speed versus cost optimization. In fast mode, the model returns results faster but uses more tokens per query due to less aggressive compression and more verbose intermediate outputs. Standard mode is cheaper but slower.
The implication: Steinberger is not optimizing for cost. He is optimizing for throughput. The three-person team wants the agents to move fast enough to keep pace with the project's development velocity. The 70% saving from switching modes is real, but it comes with a speed penalty that may reduce the agents' net contribution to the team's output.
Whether the current spend represents good value depends entirely on what the team is shipping. OpenClaw is a real open-source project with active development. If 100 agents are genuinely accelerating what three engineers can accomplish — not just generating noise — the economics might be defensible at research-pricing. The harder question is whether those same economics work at market rates, without an employer covering the difference.
What $1.3 Million Actually Buys
The token math is illuminating. 603 billion tokens at OpenAI's API pricing for GPT-5.5, even at research-discounted rates, is substantial. But the unit economics of AI-assisted development remain genuinely opaque. We know what the inputs cost. We rarely know what the outputs are worth.
Steinberger's own framing: he is building a proof point for AI-native software development. What does a team that treats AI agents as first-class infrastructure actually ship? The OpenClaw repository is the artifact. The $1.3 million per month is the investment. Whether the output justifies it is the experiment.
For the broader ecosystem, the data point is useful even if the conditions are not generalizable. Three people running 100 agents is not a typical startup configuration. It is a lab configuration. The question for readers building real businesses is: at what scale, and at what cost structure, does this math start to work?
The Subsidy Problem
AI API pricing is subsidized. This is not a secret, but it is easy to forget when looking at a viral bill screenshot. Labs are burning capital to acquire usage patterns, lock in developers, and drive distribution before commodity pricing settles. The real cost of compute behind 603 billion tokens — unsubsidized, at market rates — would be higher.
This creates a peculiar situation for anyone trying to evaluate AI agent ROI. The published prices are not the true marginal costs. The true marginal costs are not publicly known. What we have is a heavily discounted proxy that tells us very little about where the actual economics land.
Steinberger's bill, in this light, is both more and less than it appears. More, because it represents a real deployment at significant scale, with real tasks, run by people who understand the tooling deeply. Less, because the price was paid by a lab with strong incentives to keep that particular developer ecosystem alive and productive.
The story here is not that $1.3 million is too much to spend on AI-assisted development. The story is that the entire industry is running on a pricing model that has not been stress-tested against actual market conditions. When the subsidies end — and they will — the math changes.
Sources: