OpenClaw creator Peter Steinberger disclosed a 30-day OpenAI API bill of about $1.305 million, a figure generated from roughly 603 billion tokens and 7.6 million requests across around 100 Codex agents run by a three-person team.

Steinberger emphasized that most of the spend supports OpenClaw development, with OpenAI tokens reportedly provided by OpenAI and not paid out-of-pocket by him.

The billing was largely driven by Codex in Fast Mode, and proponents note that disabling Fast Mode could cut the monthly cost to around $300,000, while traditional pricing would run higher.

Analysts say the episode highlights a broader economics question: the cost gap between subscription-based AI tools and actual compute usage by autonomous agents, underscoring token management and security concerns at scale.

Some critics labeled the spend irresponsible or marketing-driven, while Steinberger framed it as an exploration of future software-building paradigms where token costs aren’t a constraint.

Tom's Hardware notes the project’s top model is listed as GPT-5.5, reflecting the cutting-edge tooling used by OpenClaw.

Steinberger’s Codex-powered agents autonomously perform tasks such as reviewing pull requests, scanning for security issues, deduplicating GitHub issues, writing fixes, monitoring benchmarks, and even generating PRs for discussed features.

OpenClaw has generated significant attention for rapid growth and open-source influence, including sparking interest in related hardware purchases and wider adoption.

Industry observers note that daily token spend hovered around $20,000, illustrating how subsidized compute can accelerate AI development and talent strategies.

Within Silicon Valley, token-usage bragging and competitive token leaderboards have emerged, with token access becoming a potential differentiator for AI efforts.

Steinberger stresses that all OpenClaw code and outputs are open source, framing the spend as research into a future of token-free software development.

OpenClaw’s majority usage came from Codex in Fast Mode; Steinberger suggests that removing Fast Mode would lower costs but keeps the project as a stress test for token-agnostic development.