Peter Steinberger spent $1.3 million on API in a month of OpenClaw experiments

Peter Steinberger, an engineer at OpenAI and creator of the OpenClaw project — a system for autonomous code generation — recently published a financial report of his month-long experiment. The numbers read like science fiction: in a single month, the bill came to $1.3 million. This is not a typo, not a hypothesis — these are real dollars spent on OpenAI API to have machines write code at industrial scale.

How the bill got so high

Over the course of 30 days, Steinberger ran approximately 100 language model instances simultaneously. Each worked in parallel, solving separate coding tasks. Over the month, 603 billion tokens were processed through 7.6 million API requests. This is not just a large number — it's a live demonstration of what happens when an AI agent starts writing code at industrial scale, without human intervention. Each request to the OpenAI API costs money. When you run 100 agents at the same time, and each can generate dozens or hundreds of requests per minute, the bill grows exponentially. This is not linear cost growth — it's a power law.

Why it's expensive: the economics of modern LLMs

• Token price is not uniform across all cases — long context (lots of text in the initial prompt) costs more
• 100 parallel sessions multiply costs exponentially, creating a scaling effect in the wrong direction
• Each request requires full context processing by the model — this is the most computationally expensive part of transformer networks
• Long-lived agents send old context in new requests, repeating payment for the same tokens
• No volume discounts in real time — OpenAI charges for each token at current rates

Steinberger suggests in his report that building a commercial product requires serious optimizations. For example: using token caching (a new OpenAI API feature that reduces the cost of repeated requests with the same context), batching — grouping multiple requests together, or using cheaper models for auxiliary tasks like code analysis. Without these tricks, scaling autonomous code generation becomes economically unfeasible.

What this reveals about compute costs

Expenses of this magnitude reveal the real, hidden economy of AI tools. When an agent is intelligent enough to work for hours without human oversight, the cost becomes visible and inevitable. This is not a criticism of OpenAI — it's the math of computational costs. Transformers are expensive to run inference on, and no optimization will change that radically.

"If you run a fully autonomous agent that works for long periods and frequently switches between models, your bill simply skyrockets," — that's roughly how

Steinberger summed up the core problem.

What will change in the market

For developers and startups, this is an important signal: AI code generation at scale requires not only technical knowledge, but also a deep understanding of compute costs. We will likely see a wave of optimizations. Companies will start looking for ways to reduce the effective price per token, use lighter models for drafts and contextual analysis, cache and reuse contexts as much as possible. For OpenAI, this may mean the emergence of new pricing plans for high-volume users — something like bulk discounts, contract rates, or special enterprise solutions. The era when AI code generation was cheap and accessible to everyone is ending. We are entering an era of thoughtful, financially optimized use of autonomous agents.