Three people with 100 AI programmers burned through $1.3 million in just one month! OpenAI: We're covering the costs.

TL;DR

• A team of three engineers spent $1,305,088.81 in 30 days using approximately 100 OpenAI Codex AI agents to handle software development tasks.
• The heavy usage resulted in 7.6 million requests and the consumption of 603 billion tokens over the one-month period.
• OpenAI is covering the cost of this specific pilot project, allowing the small team to outspend a large corporate engineering department's operational expenses.
• The AI agents performed a wide range of duties including reviewing pull requests, identifying security vulnerabilities, deduplicating issues, fixing bugs, and monitoring benchmarks.
• Peter Steinberger, the creator of the tool, noted that even with high costs, the productivity gain is significant, and costs drop below the salary of a single engineer when "fast mode" is disabled.

Why It Matters

This case study demonstrates the shift from using AI as a coding assistant to deploying it as an autonomous engineering workforce. It highlights that token economics are becoming a primary driver of software development strategy, where marginal cost of computation is replacing labor cost as the key budget variable for high-velocity teams.

Key Data

• Monthly Spend: $1,305,088.81 for the 3-person team over 30 days.
• Token Consumption: 603 billion (6030亿) tokens processed in the 30-day window.
• Request Volume: 7.6 million (760万) API requests initiated during the month.
• Team Composition: 3 human engineers managing approximately 100 concurrent Codex instances.
• Cost Comparison: Steinberger stated that without "fast mode," the operational cost falls below the salary of one standard engineer.

Technical Details

• Agent Orchestration: The team deploys around 100 Codex instances running in the cloud, specifically tasked with high-friction engineering workflows such as PR reviews, security scanning, and bug triage.
• Observability Tooling: The team utilized a custom macOS menu bar application called "CodexBar" to track usage windows, credit balances, and costs across multiple AI providers, including Codex, Claude, Cursor, Gemini, and Copilot.
• Workflow Integration: The AI agents are integrated into the development lifecycle beyond just code generation, including listening to meeting notes to autonomously open pull requests and sending regression alerts to Discord channels.
• Cost Optimization: The high expenditure is attributed to the use of "fast mode." Steinberger indicates that disabling this feature significantly reduces token costs while maintaining productivity gains, suggesting a trade-off between latency/throughput and expense.

Industry Insight

The "if" scenario of unlimited token cost is rapidly becoming a "when" scenario due to decreasing model pricing; as costs drop by orders of magnitude, small teams will be able to replicate this "three-person, 100-agent" model, fundamentally altering startup lean startup methodologies. Software engineering roles will evolve from writing code to supervising autonomous agent swarms, making "agent operations" and "token budget management" critical new skill sets for engineers.

FAQ

Q: Who is paying for the $1.3 million monthly bill?
A: OpenAI is reimbursing Peter Steinberger's team for the costs associated with this specific experimental setup.

Q: What is the role of the 100 AI agents in the development process?
A: The agents act as a digital workforce handling repetitive and complex tasks such as code review, security auditing, bug fixing, benchmark monitoring, and automating pull request creation from meeting insights.

Q: What tool was used to monitor the AI usage and costs?
A: The team used "CodexBar," a custom-built macOS menu bar tool that tracks credits, costs, and usage windows for various AI coding services.