Cloudflare event highlights AI cost concerns

At the Cloudflare Connect conference in San Francisco on October 21, Dane Knecht, Cloudflare’s technical director, mentioned that Peter Steinberger, founder of PSPDFKit and creator of OpenClaw, stated he had a "nearly unlimited token budget." Steinberger estimated the cost for the previous month at around 10 million dollars (nearly 9 million euros), according to The Register.

OpenClaw’s high costs and OpenAI’s pricing

Steinberger, who joined OpenAI in February, is the creator of OpenClaw, an open-source agent that acts on behalf of users on their computers and messaging platforms. The 10 million dollars is likely calculated at the public API rate, which is what a customer would pay, while the actual cost to OpenAI on its own servers is probably much lower.

Comparing costs over time

In May, Steinberger shared a screenshot on X of his tracking tool, CodexBar, which showed 1.3 million dollars over 30 days (approximately 1.16 million euros). When asked if OpenAI was charging him for this, he replied, "of course not," indicating that the bill was multiplied by nearly eight in four months.

Scale of usage and implications

The 10 million dollars from September represents approximately 330,000 dollars (nearly 295,000 euros) per day, or the equivalent of 500,000 ChatGPT Plus subscriptions at 20 dollars per month. In May, the record showed 603 billion tokens and 7.6 million requests in 30 days, mostly on GPT-5.5. Steinberger was also using "Fast Mode," which is more expensive, but it is unclear whether the September costs included this mode, making the comparison with May fragile.

OpenClaw’s goals and performance

Steinberger explained on X in May that he is running "about 100 Codex in the cloud" continuously. These agents re-read every contribution, look for security flaws, eliminate duplicates and spam, and even listen to team meetings to start work while they are discussed. His goal is to discover how software will be built "when tokens no longer matter."

Results and limitations of the project

The result is visible in the OpenClaw 2.0 post, published on August 30, which shows 933 contributors and more than 16,000 changes integrated, approximately half of all changes the project has ever accepted. The project remained nearly seven weeks without publishing a version, whereas it had released 106 in 230 days previously. However, this version 2.0 was released before September, and it shows what the machine produces, but nothing of what the 10 million from September produced.

Implications of high usage

At this rate, one can assume that a large part of the agents' work consists of redoing, correcting, and restarting, which would explain Steinberger's comment that "not everything is useful." Moreover, nothing indicates that the 10 million per month are solely for OpenClaw. It is possible that the developer also helps with the upgrade of GPT and Codex at OpenAI with these tokens.

Infrastructure and local vs. cloud computing

Steinberger lamented that the infrastructure does not follow, stating, "Why is it still using my computer? All of this should not be in the cloud?" The Register raises the opposite question: shouldn't it rather run locally? An agent like OpenClaw reads your emails, messages, and files. The OpenClaw 2.0 post itself cites the example of an agent that monitors the emails of a child's school, the kind of data one would prefer to keep at home.

Performance of local models

The project also holds a different discourse than its creator: during installation, OpenClaw 2.0 detects local models already present on the machine, just like ChatGPT or Claude subscriptions and API keys. Testing on a mini-PC with the Qwen3 Coder 30B model, which the Ollama software runs within OpenClaw, showed that the first request, which sends about 20,000 tokens at once, takes 3 to 4 minutes. The following ones drop to 4 to 8 seconds as long as the model remains loaded in memory. It works, but a local model does not do the work of a cutting-edge model.

Trust and agent behavior

Regarding the trust to grant an agent, Steinberger acknowledges that "it is a difficult question." He advises not to authorize everything and to delegate permissions to a second agent, which grants or denies and "negotiates" with the first. He admits, however, that a large part of the behavior depends on the model itself, and thus on the trust one grants it. According to him, the monitored agent often reconsiders itself or finds a simpler solution that remains within the boundaries.