Anthropic over the weekend told subscribers they'd have to pay up for heavy use of its Claude AI models to power third-party agents like OpenClaw. 

Users with monthly subscriptions can still use Claude models, including Opus, Sonnet and Haiku, through these third-party agents. But you'll have to pay via Anthropic's API or use a "pay-as-you-go option" that will be billed separately from your Claude subscription payment.

"The $20/month all-you-can-eat buffet just closed," wrote AI product manager Aakash Gupta on X. 

At the same time, Anthropic recently announced new features that bring some of the things that made OpenClaw so popular into Claude itself. Claude can use your computer, even if you're not at it, for example. 

Why this policy matters 

There has been growing tension between OpenAI and Anthropic, recently inflamed by the controversies involving contracts with the US Defense Department. But there is also tension between users who want to run autonomous AI agents constantly and the AI labs that are trying to control costs by managing the tasks their models are used for. 

Claude is a chatbot that was created to be prompted by humans, not for millions of AI agents to use it for workflows. These agent tools, like Manis and OpenClaw, require much more power to run and burn through tokens faster than regular human chatting. Anthropic has already taken steps to address the demand that heavy agent users bring, like a five-hour session cap during peak periods for the models.

"We've been working to manage demand across the board, but these tools put an outsized strain on our systems," Anthropic wrote in its email to customers.

OpenAI has been all-in on agentic tools. Early this year, the AI company hired Peter Steinberger, the creator of OpenClaw, with the aim of bringing AI agents to a broad audience. Steinberger was vocal about his critiques of Anthropic's new policy, taking to X over the weekend. 

"Funny how timings match up, first they copy some popular features into their closed harness, then they lock out open source," he wrote. 

(Disclosure: Ziff Davis, CNET's parent company, in 2025 filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

The future of agent compute power 

Friction between heavy agent users and AI companies is likely to get worse. These AI agent tools are extremely powerful: they can run for hours, take actions across apps like Gmail, Slack and iMessage, and work autonomously much longer and faster than a human could. Because of this power, they are far more costly and require far more power to run compared to a human prompting a bot. It's likely that AI companies will increasingly push these compute costs onto heavy users through price increases or steps like those taken by Anthropic.