100 AI Agents, $1.3M/Month: What OpenClaw's Tab Reveals

Peter Steinberger’s team spent $1.3 million on OpenAI API tokens in a single month running 100 concurrent Codex agents on a three-person codebase. That’s 603 billion tokens and 7.6 million API requests — and the bill is the most detailed public data point yet on what agentic engineering actually costs when the constraints come off.

The Economics of Always-On Agents

OpenClaw is an open-source developer tooling project. Steinberger’s setup isn’t a controlled experiment — it’s a production workload where AI agents are the primary engineering workforce. The 100 Codex instances run continuously: reviewing pull requests, scanning commits for security vulnerabilities, deduplicating GitHub issues, writing fixes, and opening new PRs based on the project’s broader roadmap. Some agents monitor benchmark regressions and post results to Discord. Some attend meetings and generate feature proposals from what they hear.

OpenAI covers the bill as part of a promotional arrangement, which is why the Fast Mode pricing — the most expensive configuration — was left on. Without it, Steinberger estimates the monthly cost drops to roughly $300,000. That $1.3M headline is a ceiling, not a floor.

What the Numbers Actually Mean

The $300K baseline is still not a small number, but it reframes the story. At that level, a team of 100 hypothetically equivalent human engineers — at mid-market US salaries plus benefits — would cost somewhere between $15M and $25M annually. Even the inflated $1.3M/month figure is $15.6M annualized, which sits near the low end of that range for a fraction of the headcount.

The comparison breaks down quickly under scrutiny — agents don’t sleep, but they also don’t make architectural decisions, handle stakeholder relationships, or transfer institutional knowledge. What Steinberger’s setup demonstrates isn’t that AI agents replace engineers; it’s that a small team can now sustain the output volume of a much larger one on specific execution tasks.

The token economics also point toward cost compression over time. At 603 billion tokens monthly, even a 10% reduction in token costs — which has been the rough trajectory for frontier models — would save $130,000 per month at the current burn rate. Teams that operationalize agent fleets today are building cost models against a declining baseline.

The Pattern

This vertical has now published multiple posts tracking agentic AI cost structures and billing splits, from the Anthropic SDK credit separation to Claude Code’s compute economics. The pattern is consistent: production-scale agentic deployments are beginning to produce the kind of real-world data that was missing from the benchmark-only conversation.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.