A USD 1.3 Million Token Tab as the New R&D Budget
When OpenClaw creator Peter Steinberger shared his CodexBar dashboard, the numbers stunned even hardened AI insiders. Over 30 days, his usage on OpenAI’s API alone tallied USD 1,305,088.81 (approx. RM6,010,000), covering 603 billion tokens across 7.6 million requests. On a single day, May 15, the bill reached USD 19,985.84 (approx. RM92,000), with 19 billion tokens and 206,000 requests, largely driven by the gpt-5.5-2026-04-23 model. That kind of AI token cost rivals the annual compensation of multiple senior engineers, yet it is being burned purely on inference and experimentation rather than salaries or traditional infrastructure. Steinberger argues he could cut the cost by 70% by switching off “fast mode,” likening it to the price of a single employee. Still, the sheer scale of this AI infrastructure expense raises fundamental questions about how enterprises will budget for large-scale AI development.
OpenAI Pricing, Subsidies, and the Hidden Cost of Compute
Steinberger’s personal bill is striking, but it does not reflect what a typical customer would actually pay. He now works at OpenAI and openly notes that his spend is effectively covered as “perks of OpenAI supporting OpenClaw,” confirming that OpenAI does not charge him for these tokens. This arrangement highlights how access to free compute has become a powerful recruiting and retention lever in the AI talent wars, and underscores how heavily AI API pricing is subsidised. Commenters pointed out that the real cost of the compute behind those 603 billion tokens would be substantially higher if fully passed through. Labs like OpenAI are absorbing massive AI infrastructure expenses to grow usage and lock in developer habits. For enterprises, this means current OpenAI pricing may understate the long-term cost of their AI roadmaps, as today’s subsidised rates cannot remain disconnected from underlying economics forever.
Tokenmaxxing, Dogfooding, and the New Enterprise AI Workflow
The episode also shines a light on emerging “tokenmaxxing” culture, where teams aggressively push AI systems to their limits. Steinberger’s usage is largely tied to OpenClaw and a constellation of side projects, from device sleep tools to systems that let AI agents make phone calls. He describes experimenting with a future in which “tokens don’t matter,” using agents that listen to meetings, spin up tasks automatically, and moderate spam. This kind of dogfooding provides OpenAI with invaluable product validation and stress testing across real developer workflows. For enterprises, it previews a world where AI agents quietly handle coordination, quality control, and routine coding tasks in the background. The catch is that such workflows are token-hungry by design. Unless AI token costs fall dramatically—or remain subsidised—only organisations with deep pockets or special access to free compute will be able to fully embrace this always-on, agent-driven model.
Can Enterprise AI Spending Scale Sustainably?
Steinberger’s dashboard crystallises a broader concern: AI spending is accelerating faster than clearly measurable value. Former tech leaders warn that cash, not energy, is the true bottleneck for AI, and Steinberger’s USD 1.3 million (approx. RM5,990,000) monthly run rate embodies that tension. CodexBar and OpenClaw are influential, but they are unlikely to generate revenue matching such a burn, especially once subsidies are stripped out. For most companies, enterprise AI spending at this level would demand either enormous productivity gains or direct revenue uplift to be rational. Otherwise, AI experimentation risks drifting into bubble territory, where distorted pricing masks fragile unit economics. The lesson for enterprises is to treat current token prices as provisional. Business leaders need to model scenarios where AI infrastructure expenses rise, and ensure their AI projects deliver tangible savings or revenue, not just impressive dashboards of tokens consumed.
Small Teams, Big Models, and the New Competitive Frontier
One of the most intriguing implications of Steinberger’s token binge is competitive, not technical. His small, “extremely lean” team uses AI agents to automate meeting follow-ups, code-related tasks, and content moderation, effectively amplifying their output to resemble that of a much larger engineering organisation. In theory, this is the promise of AI: small teams wielding large models to punch far above their weight. In practice, the OpenClaw story shows this only works if someone else underwrites the AI token costs. Enterprises without subsidised access must weigh whether similar automation justifies a recurring token bill that could rival headcount. Over time, we may see a divide: companies with free or discounted compute building aggressive AI-first workflows, and everyone else adopting AI more cautiously. The future of enterprise AI competitiveness may hinge less on talent alone and more on who can afford to keep the tokens flowing.