A $1.3 Million Token Bill That Looks Like a Team’s Salary
When developer Peter Steinberger shared a screenshot of his CodexBar usage dashboard, the numbers stunned even hardened AI insiders. Over 30 days, he had consumed 603 billion tokens across 7.6 million requests, translating to USD 1,305,088.81 (approx. RM6,008,000) in OpenAI API spending. On a single day, his bill hit USD 19,985.84 (approx. RM92,000), driven largely by the gpt-5.5-2026-04-23 model. Commenters quickly pointed out that this is equivalent to the annual cost of multiple senior engineers, but spent purely on AI token costs rather than salaries, infrastructure, or equity. Steinberger’s defense was revealing: by disabling “fast mode,” he said, the same work would be about 70% cheaper, or roughly comparable to one employee. His experiment crystallizes a new reality: AI development expenses can now rival human payroll, and for some teams, compute has become the primary line item.
OpenAI-Funded Compute and the New Talent Perk
Steinberger isn’t personally paying these AI development expenses. Since joining OpenAI, he says the massive OpenAI API spending behind OpenClaw is covered as “perks of OpenAI supporting OpenClaw,” confirming in replies that he is not charged for tokens. This arrangement highlights a major shift: access to effectively free AI compute is emerging as a core benefit in the AI talent wars, similar to how stock options or cloud credits once attracted engineers. Inside labs, token leaderboards and “tokenmaxxing” culture turn high usage into a badge of honor, encouraging developers to push AI agent productivity to its limits. But such subsidized access also distorts market signals. When the true cost of AI compute scaling is masked, engineers can explore aggressive, AI-first workflows that would be hard to justify for independent developers or cash‑constrained startups paying retail prices.
How Work Changes When Tokens ‘Don’t Matter’
Steinberger says his goal is to explore “how we would build software in the future if tokens don’t matter.” In practice, that means AI agents saturate his workflow: they listen to meetings and start working based on conversation alone, review comments for spam, and support a flurry of projects ranging from device sleep tools to systems that let AI agents place phone calls. This is AI-first development taken to an extreme, where human engineers orchestrate workflows while agents handle much of the execution. Steinberger argues this lets his team run “extremely lean,” swapping headcount for compute. Critics counter that any project burning USD 1.3 million (approx. RM5,980,000) in tokens per month can hardly be called lean. The tension underscores a key question: at what point does substituting engineers with AI agents deliver real productivity gains rather than just spectacular bills?
The Accessibility Gap: Who Can Actually Build This Way?
Steinberger’s token usage is not just a curiosity; it exposes an emerging divide in AI software engineering. Large labs and well‑funded companies can treat high AI token costs as strategic investment, absorbing enormous AI development expenses in pursuit of future dominance. Independent developers and smaller teams, however, face a different calculus. Without corporate backing, they cannot casually iterate through billions of tokens to explore every possible AI agent workflow. This raises practical questions about accessibility and fairness: if the most advanced AI-first methodologies require subsidized compute at scale, they risk becoming the privilege of a few. Teams without these subsidies must optimize every prompt, carefully monitor OpenAI API spending, and limit experimentation. As AI compute scaling accelerates, the industry must decide whether the future of coding is open to all or gated behind massive, invisible token budgets.
Are We Seeing Sustainable Productivity or a Subsidized Bubble?
Supporters of aggressive AI spending point to clear success stories, such as AI compressing months of specialized research work into days, offering obvious productivity gains over expensive human labor. In contrast, Steinberger’s spending is harder to map directly to revenue: CodexBar is useful and OpenClaw has grown rapidly, but neither evidently generates USD 1.3 million (approx. RM5,980,000) per month. Critics argue this gap between token consumption and measurable value hints at bubble dynamics, fueled by heavily subsidized AI token costs. Labs are racing to lock in usage, shouldering losses today in hopes of scale tomorrow. The result is a market where developers normalize burning through sums that could fund entire teams. Whether this heralds a new, hyper‑productive era of AI agent–driven development, or simply a phase of distorted pricing on the way to a correction, remains an open—and increasingly expensive—question.