"Not all of it was useful": OpenClaw developer spent $10 million on AI tokens in a month
Peter Steinberger, an OpenAI employee and creator of OpenClaw, said that last month he spent about $10 million on tokens to run AI tasks.
Peter Steinberger, an OpenAI employee and creator of OpenClaw, said that last month he spent about $10 million on tokens to run AI tasks.
This frank statement was made during the Cloudflare Connect 2026 conference in San Francisco in a conversation with Cloudflare CTO Dane Knecht, The Register writes .
When the audience responded with a surprised roar, Steinberger added, “People don’t always understand that not all of this was actually useful.” However, the financial costs do not fall into the developer’s pocket — all expenses are covered by OpenAI’s investor funds.
Clouds instead of local machines and the open source crisis
During the discussion, Steinberger expressed surprise that most computing is still tied to personal computers, believing that all AI processes should move entirely to the cloud. However, Dane Knecht reminded us of the financial and privacy constraints for ordinary developers who do not have unlimited venture budgets.
The speakers also touched on the problem of the decline of open source due to the influx of AI-generated code. Steinberger noted that the approach to reviewing pull requests needs to change: “If I can see the flow of your words — how exactly you got from point A to point B — my confidence in the result is much higher than if you just throw ready-made code over the fence.”
Instead of hundreds of personal AI agents for each user, the creator of OpenClaw predicts the emergence of a single background intelligence that will anticipate a person's needs and provide answers before they even ask a question.
Control and security: an agent who checks the agent
The biggest challenge for deploying autonomous systems remains trust and granting access rights. According to Steinberger, the behavior of AI depends primarily on the underlying model itself, not on the external code wrapper.
To minimize risks (e.g., prompt injection), the developer suggests not granting all permissions to one agent, but using a two-agent architecture.
“You need another agent to evaluate and grant permissions to the first one. They have to negotiate with each other. In many cases, the agents themselves come to the conclusion: ‘I probably shouldn’t do this’ or ‘There’s an easier way within the established rules.’ Combining the two of these systems significantly increases the level of trust in autonomous AI.”
Anthropic temporarily bans OpenClaw creator from accessing Claude
Anthropic temporarily bans OpenClaw creator from accessing Claude
The developer of OpenClaw — autonomous AI assistants for PCs — has joined OpenAI. What is the fate of his project?
The developer of OpenClaw — autonomous AI assistants for PCs — has joined OpenAI. What is the fate of his project?
Read the country's main IT news in our Telegram
Read the country's main IT news in our Telegram
