austrian developer peter steinberger recently unveiled a dashboard displaying openai api usage data, revealing that his team incurred $1.3 million (approximately rmb 8.9 million) in expenses over 30 days. this cost covered 7.6 million api requests, totaling 603 billion tokens, all processed by roughly 100 codex agent instances—supported by a three-person technical team. notably, steinberger himself joined openai in february this year, with all associated resource consumption borne by the company. on the very day the screenshot was published, daily spending approached $20,000, handling over 200,000 requests.
as the founder of the open-source project openclaw, steinberger systematically explores the boundaries of ai-native software development under the premise of “zero budget constraints.” the codex agent cluster he deployed boasts end-to-end autonomous collaboration capabilities: automatically reviewing pull requests, identifying security vulnerabilities, deduplicating and categorizing tickets, generating fix code and initiating merges, monitoring service performance metrics in real time, and even integrating into online meetings to understand discussions and automatically draft feature requirement proposals. since its launch, the project has continuously stirred industry-wide excitement—leading, for instance, to meta’s head of ai alignment receiving their inbox flooded with automated reports, and indirectly prompting nvidia to accelerate the initiation of similar ai engineering agent research and development programs.
steinberger pointed out that the $1.3 million bill resulted from peak billing rates when codex operated in “turbo mode”; switching to standard mode could reduce costs to around $300,000. compared with openai’s publicly disclosed average monthly expenses for typical developers ($100–$200), his usage has already surpassed the theoretical upper limit of the cost model. in response, he admitted he is not concerned: “this isn’t an operating expense—it’s a controlled experiment.” the core objective is to verify what structural shifts will occur in software development paradigms once computational power, token costs, and labor expenses cease to be bottlenecks. all technical implementations, toolchains, and evaluation methodologies are fully open-sourced under the mit license.
