A Single Developer, 603 Billion Tokens, and a Multi-Million Bill
Peter Steinberger, creator of the AI agent platform OpenClaw, recently revealed a dashboard that stunned the developer community: in 30 days, his projects consumed 603 billion tokens across 7.6 million requests, translating into USD 1,305,088.81 (approx. RM6,008,000) in AI token costs. On one peak day alone, May 15, his usage hit 19 billion tokens, 206,000 requests, and nearly USD 19,985.84 (approx. RM92,000) in spend, largely powered by OpenAI’s gpt-5.5-2026-04-23 model. For comparison, commentators noted that such a monthly API spending level is similar to what several senior engineers might earn in a year—yet here it is being burned purely on compute. Steinberger’s CodexBar menu bar app tracks usage across multiple AI providers, but this particular bill highlights OpenAI API spending at a scale that most startups or independent developers could never sustain out of pocket.
Inside Token Billing Models and the New Cost of Building Software
Behind the eye-popping numbers is a simple mechanism: token billing models. Instead of charging flat subscription fees, leading AI companies price access per token—the tiny units that represent chunks of text processed by models. Steinberger’s CodexBar screenshot shows how this can balloon as developers run millions of requests during rapid prototyping, testing agent workflows, and instrumenting everyday tasks like summarizing meetings or triaging spam. At this scale, API spending becomes a core line item in AI development expenses, not a marginal cost. Steinberger even notes that merely switching off a “fast mode” could make usage around 70% cheaper, illustrating how latency, model choice, and configuration directly translate into cash burn. This is modern software R&D: instead of buying more servers or hiring more engineers, teams increasingly buy more tokens—essentially renting intelligence and compute on demand.
Subsidized Pricing, Talent Perks, and the Question of Sustainability
The controversy around Steinberger’s figures is not only about scale, but who pays. He now works at OpenAI and has clarified that the tokens powering OpenClaw are covered as a perk—“ofc not,” he replied when asked if he was personally billed, calling it “OpenAI supporting OpenClaw.” Critics argue that this is a vivid example of heavily subsidized AI token costs: labs are absorbing enormous compute expenses to attract users, capture market share, and lure talent with free or discounted API access. Former executives and investors have warned that the real constraint on AI is cash, not energy, and this kind of spending underscores that point. If labs passed on the full, unsubsidized cost of compute, many experiments would become economically impossible. The current ecosystem is therefore built on aggressive discounting and investor capital, raising doubts about how sustainable today’s pricing truly is.
Accessibility for Smaller Developers in a Tokenmaxxing Era
Steinberger’s usage sits within a broader “tokenmaxxing” culture in which power users compete on leaderboards for who can push the most tokens through cutting-edge models. For engineers with corporate backing or lab sponsorship, high token bills become a badge of honor and a hiring perk—free compute is part of the compensation package. For independent developers and small teams, though, such AI development expenses pose a serious barrier to entry. When one engineer’s monthly usage could “bankroll a small startup,” as one commenter put it, many others are effectively priced out of large-scale experimentation. This bifurcation risks concentrating innovation among those with privileged access to cheap or free tokens. As token billing models become the default way AI is monetized, the industry must confront whether its most powerful tools remain accessible only to insiders, or whether pricing evolves to support a broader developer ecosystem.
From Experimental Burn to Viable Business Models
Defenders of massive API spending argue that it is analogous to early cloud or search investments—an upfront burn for future advantage. Steinberger has framed his approach as exploring “how we would build software in the future if tokens don’t matter,” using AI agents to listen to meetings, launch work automatically, and handle moderation tasks so that OpenClaw can operate “extremely lean.” Yet many observers question whether a project incurring USD 1.3 million (approx. RM5,980,000) in token usage per month can be called lean at all, especially if its revenue falls far short of that figure. The gap between token consumption and clear value creation lies at the heart of current AI economics. For token-based business models to endure, AI output must reliably generate more value than it costs in compute—whether via productivity gains, new products, or defensible strategic moats. Until then, million-dollar token bills will continue to look less like efficient R&D and more like speculative bets.