A USD 1.3 Million Token Month, Explained
When OpenClaw creator Peter Steinberger shared a screenshot of his CodexBar dashboard, the numbers stunned even seasoned AI watchers. Over 30 days, his OpenAI API usage showed a spend of USD 1,305,088.81 (approx. RM6,020,000), covering 603 billion tokens across 7.6 million requests. On a single day, 15 May, the dashboard logged USD 19,985.84 (approx. RM92,000), 19 billion tokens and 206,000 requests, with gpt-5.5-2026-04-23 leading the charge. Those AI token costs are roughly comparable to the annual compensation of multiple senior engineers in a major tech hub, but here the money is going purely into compute. The episode has become a vivid case study in large language model pricing: what happens when an individual developer, empowered by access to frontier models, runs AI tooling at a scale previously reserved for big companies’ backends.
Subsidised Compute and OpenAI’s Quiet Talent Perk
Steinberger’s colossal AI token bill isn’t hitting his personal bank account. Since he now works at OpenAI, he confirmed that the company doesn’t charge him for these tokens, calling the spend a perk of OpenAI supporting OpenClaw. In an era when AI talent is fiercely contested, free or heavily subsidised compute has become a key weapon in hiring and retention. OpenAI spending on internal token usage mirrors a broader trend: competitive token leaderboards, “tokenmaxxing” culture and internal races to push models to their limits. For employees, this effectively makes AI infrastructure expenses vanish at the point of use. Once tokens feel free, developers can explore architectures and workflows that would be economically irrational at retail prices, blurring the line between serious R&D, personal experimentation and marketing spectacle.
What Running 100+ AI Agents Really Costs
Behind the sticker shock is a specific style of development: a dense swarm of AI agents running in parallel. Steinberger described AI agents that listen to meetings and start working based on spoken instructions, review comments for spam and automate a long list of small but persistent tasks. Building OpenClaw and dozens of related projects means orchestrating 100+ AI agents that constantly call large language models, driving up token counts and, in turn, AI infrastructure expenses. He noted that disabling “fast mode” could make his workload roughly 70% cheaper, arguing his effective spend could be compared to a single employee. Yet the volume remains extraordinary, demonstrating how modern workflows—code generation, autonomous task routing, continuous monitoring—translate directly into massive AI token costs when scaled across millions of requests.
When Tokens Don’t Matter, Behaviour Changes
Steinberger said he is explicitly exploring a future where “tokens don’t matter”: how software might be built if AI usage felt effectively free. That mindset, enabled by OpenAI’s subsidies, changes how developers reason about cost–benefit trade-offs. Tasks previously done manually or avoided entirely—full-meeting transcription, continuous code review, agent-driven automation for niche workflows—suddenly become default behaviours. For many startups or independent developers, such large language model pricing is prohibitive; they must constantly weigh each call against burn rate. But when OpenAI spending absorbs the bill, token costs become psychologically negligible. The result is a wave of experimentation that pushes the frontier of what’s technically possible, while simultaneously masking the true economic weight of running this kind of always-on, agent-heavy infrastructure.
A Glimpse Into AI’s Bubble Question
The reaction on X ranged from awe to open scepticism. Commenters noted that Steinberger was effectively “burning through enough tokens to bankroll a small startup,” while others questioned whether those funds would be better deployed hiring human engineers. Critics argued that AI API pricing is heavily subsidised and that the real, unsubsidised compute cost behind 603 billion tokens would be far higher. This disconnect between apparent usage and visible output fuels the debate over whether current AI economics resemble a bubble. Capital-rich labs absorb losses to dominate distribution, while developers explore workflows that may not yet produce proportional value. Steinberger’s usage shows both the power and expense of AI-driven development: it can make a tiny team feel enormous, but only because someone else is quietly picking up an enormous infrastructure tab.