according to gizevo technology, austrian developer and founder of the open-source project openclaw, peter steinberger, recently posted a screenshot of an openai api bill on social media, sparking widespread attention within the industry. the data shows that his team spent as much as $1.305 million on api fees over the past 30 days—equivalent to 603 billion token calls and 7.6 million requests—all executed by roughly 100 codex instances maintained by a core team of three, dedicated to serving openclaw’s cutting-edge ai-assisted development experimental platform.
steinberger officially joined openai in february this year, with all related expenses fully covered by the company. the bill reveals that gpt-5.5 was the most frequently used model; on the day the screenshot was published, his account incurred $19,985.84 in spending and made 206,000 api calls.
the codex agent cluster built by the team has achieved a high degree of autonomy: it can automatically review pull requests, identify security risks in code submissions, deduplicate github issues, and generate fix patches. some agents proactively create prs based on project roadmaps, while others continuously monitor performance benchmarks and instantly push alerts to the team’s discord channel whenever regression issues are detected. additionally, certain agents can even access conference audio streams, comprehend discussion content, and generate feature-level code proposals accordingly.
openclaw has frequently made headlines lately, drawing public attention for incidents such as emptying the inbox of meta ai’s alignment director and pressuring nvidia to accelerate the release of competing products. steinberger consistently describes it as an “ai engineering stress test with no budgetary limits,” emphasizing that its core value lies in rigorously validating the feasibility and boundaries of ai-native development paradigms.
in a subsequent clarification, he explained that the $1.3 million bill resulted from activating codex’s “fast mode”—a setting that consumes credits at extremely high speed, far exceeding normal operational rates. if this mode is disabled, the original api costs could be reduced to approximately $300,000. even so, this figure remains highly significant: the codex pro subscription plan costs only $200 per month yet provides api usage equivalent to $5,000–$6,000; translated, under non-fast mode, the actual consumption would be comparable to running about 60 pro accounts simultaneously.
openai officially estimates that codex costs each developer roughly $100–$200 per month on average, but explicitly notes that this range varies significantly depending on model selection, automation depth, and task complexity. while the steinberger team’s usage represents an extreme case, it precisely highlights the structural gap between end-user pricing and the true underlying computational costs.
currently, ai programming tools are collectively entering a phase of cost rationality. mainstream platforms like codex, claude code, and cursor, while vying for developers’ attention, generally employ subsidy strategies that keep prices below the official api list price. meanwhile, since april, openai has transitioned codex entirely to a token-based billing system, enhancing pricing transparency while amplifying cost volatility awareness among high‑usage users—behind the efficiency gains, the ledger of computational expenses is becoming increasingly clear and incisive.
