Peter Steinberger's three-person team runs approximately 100 AI agents powered by OpenAI's Codex model for the open-source OpenClaw project, incurring $1.3 million monthly in API costs. Steinberger treats this spending as a research investment to understand what software development looks like when token expenses become irrelevant.
The agents perform multiple development tasks: writing code, reviewing pull requests, and identifying bugs. This scale of deployment represents one of the largest known continuous uses of GPT-based coding models in production workflows.
Steinberger's approach diverges from the typical cost-conscious optimization most organizations pursue when deploying AI. Rather than minimizing API spend, he maximizes agent instances to explore the boundaries of what's possible when computational constraints disappear. The experiment generates insights into agent behavior, development velocity, and quality outcomes that wouldn't emerge from cost-limited deployments.
The $1.3 million monthly bill reflects both raw token consumption and the computational intensity required to maintain 100 concurrent agents. Each agent operates independently, processing code, analyzing diffs, and running debugging operations continuously.
This experiment highlights a growing trend where well-funded teams leverage abundant compute to advance AI capabilities research. Unlike commercial optimization, which prioritizes efficiency, Steinberger's model prioritizes discovery. The data generated from unrestricted agent operations helps inform understanding of AI-assisted development at scale.
OpenClaw's infrastructure demonstrates the shifting economics of AI development. What remains unclear is whether insights from unlimited-budget scenarios translate to practical improvements for cost-conscious organizations. The project also raises questions about optimal agent architecture, token efficiency, and whether current models reach capability plateaus regardless of compute allocation.