Spending $1.3 M on Tokens in One Month: Peter Steinberger Reveals OpenAI‑Covered Bill

Peter Steinberger, known as the "father of lobster," revealed that his AI‑driven project burned through 603 billion tokens in 30 days, generating roughly 7.6 million requests and a total cost exceeding $1.3 million. He noted that disabling the "fast mode" would have cut the price by about 70%, bringing the expense down to roughly the cost of a single employee.

In response to criticism, Steinberger explained that the entire codebase was generated by OpenAI Codex, while many of the messy pull‑requests were later cleaned up by Claude. He described a highly automated workflow that runs about 100 Codex instances in the cloud to review every pull request and issue, automatically close stale issues, and generate security reviews.

The automation includes agents that can reproduce complex environments, launch temporary "crabbox.sh" machines, record video demos, and post before‑and‑after comparisons in PRs. Additional Codex agents monitor new issues, create PRs when they align with the product vision, and even scan comments for spam, banning offending users. Performance benchmarks are continuously validated, with regression reports sent to Discord.

Steinberger’s high token consumption sparked a broader discussion about "Tokenmaxxing," a trend where companies like Meta and Amazon publicly rank internal token usage and treat token throughput as a KPI. Meta’s top individual users averaged 281 billion tokens, translating to multi‑million‑dollar expenses depending on model pricing. Former Tesla and OpenAI scientist Andrej Karpathy has also warned that token throughput is becoming a critical metric for AI productivity.

OpenAI CEO Greg Brockman recently tweeted that tokens are rapidly becoming a universal input for problem solving, reinforcing the view that tokens are evolving into a new production resource. However, Steinberger cautions that sheer token volume is not the sole success factor; a well‑structured project management approach and reliable verification loops are essential for achieving high engineering density with small teams.

Ultimately, the article raises the question of who bears such massive token costs, noting that OpenAI does not charge Steinberger for his usage, and invites readers to consider the long‑term viability of token‑driven AI development models.