A software developer sets up an AI agent to manage email, schedule meetings and monitor a shared inbox. It works nonstop, firing off requests through the night, completing tasks no human would stay awake to manage. By morning, it has consumed more compute than most users generate in a month. Multiply that pattern across thousands of subscribers running the same setup, and suddenly a $20 flat monthly fee no longer makes sense.
That’s the reality artificial intelligence company Anthropic confronted when it ended Claude subscription access for third-party agent frameworks. Effective April 4, Claude Pro and Max subscribers can no longer use their plan limits to power tools like OpenClaw, as reported by VentureBeat.
Users who want to keep those agents running must now switch to pay-as-you-go usage bundles or connect through a direct API key. Boris Cherny, Anthropic’s head of Claude Code, announced the change on X, saying subscriptions were never designed for the kind of continuous, automated demand these tools generate. Affected subscribers received a one-time credit equal to their monthly plan cost.
When Agents Run Continuously, Caps Stop Adding Up
It comes down to a simple mismatch: Consumer subscriptions were built around human behavior. AI agents don’t behave like humans. A person typing prompts into a chat interface generates a limited number of requests, constrained by time, attention and sleep. Agents don’t stop. They run in loops, execute tasks in parallel, and generate a constant stream of activity that can quickly outpace any individual user.
Before changing access rules, Anthropic had already started to feel that strain, and had introduced stricter session limits during peak business hours, a change the company said would affect up to 7% of users, as reported by PYMNTS. The OpenClaw decision pushes these controls further, extending them to how access is granted.
The numbers tell the story. A $200-per-month Claude Max subscription was being used to run anywhere from $1,000 to $5,000 worth of agent compute tasks, according to The Decoder. Anthropic’s own tools reduce costs by reusing previously processed context. Third-party frameworks like OpenClaw skip that efficiency layer entirely, triggering full-cost processing on every request.
The new policy applies to all third-party frameworks and will expand beyond OpenClaw. Users can still use Claude models with those tools, but subscriptions will no longer cover the compute.
A Pattern Taking Shape Across the Industry
Anthropic isn’t alone in drawing a line.
Google now spells out exactly what users get: Gemini’s 2.5 Pro plan limits free users to 5 prompts per day, while $20-per-month AI Pro subscribers receive 100, and $250 AI Ultra subscribers get 500, the PYMNTs reporting shows. The move from vague tiers to hard caps reflects how standard usage controls have become across major model providers. Cursor and Replit both revisited pricing in 2025 to curb disproportionate usage, resetting expectations around what subscriptions deliver.
At the same time, demand for agentic AI is accelerating. In August 2025, 52% of companies surveyed said they were still exploring the technology, according to PYMNTS Intelligence. By November, that number dropped to 30%, with nearly 1 in 4 chief product officers reporting active pilots or full production deployments.
Developer Backlash and the Open-Source Question
The backlash was immediate. OpenClaw creator Peter Steinberger, who joined OpenAI in February, said he and fellow investor Dave Morin attempted to negotiate with Anthropic directly and received only a one-week delay, as reported by The Decoder. He accused Anthropic of incorporating OpenClaw features into its own closed system before locking out open-source alternatives, pointing to the recent addition of Discord and Telegram messaging to Claude Code.
One developer said switching API rates would make continued use cost-prohibitive and that the change would likely push users toward competing models, as covered by VentureBeat. For now, alternatives exist. OpenAI and several Chinese AI companies still support OpenClaw-style usage under subscription-like access, according to TNW. The open question is how long that model holds as agent demand and infrastructure pressure continues to climb.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.