Anthropic Ends Unlimited Claude Access for Third‑Party AI Agents, Shifts Heavy Users to Pay‑As‑You‑Go
Key Points
- Anthropic ends the unlimited $20/month plan for heavy Claude usage via third‑party agents.
- Users must now pay through Anthropic’s API or a pay‑as‑you‑go option for extensive agent workloads.
- New Claude features let the model run tasks on a user’s computer even when they’re away.
- A five‑hour session cap during peak times was already in place to manage demand.
- OpenAI hired OpenClaw creator Peter Steinberger, who publicly criticized Anthropic’s policy.
- Industry analysts expect more cost‑shifting measures as AI agents become more powerful.
Anthropic announced this weekend that its $20‑per‑month all‑you‑can‑eat plan for Claude will no longer cover heavy usage through third‑party agents such as OpenClaw. Subscribers can still access Claude models, including Opus, Sonnet and Haiku, but any extensive use via external tools will be billed separately through Anthropic’s API or a pay‑as‑you‑go option. The move follows growing pressure on AI labs to curb token‑heavy workloads and comes as the company rolls out new features that embed popular agent capabilities directly into Claude.
Anthropic told its Claude subscribers over the weekend that the unlimited access tier that allowed heavy use of its language models through third‑party agents is being discontinued. The $20‑per‑month "all‑you‑can‑eat" plan, which many developers used to power autonomous tools like OpenClaw, will now require separate billing for any substantial token consumption via Anthropic’s API or a pay‑as‑you‑go option.
According to a post on X by AI product manager Aakash Gupta, the unlimited buffet is officially closed. Users who maintain a monthly Claude subscription can still operate the model through agents, but the cost of those sessions will be tracked and charged apart from the base subscription fee.
Anthropic’s decision arrives as the company unveils new functionality that brings some of OpenClaw’s most popular features straight into Claude itself. The updated Claude can now execute tasks on a user’s computer even when the user is away, effectively narrowing the gap between the closed‑harness model and open‑source agent tools.
Industry observers note that the policy shift reflects a broader tension between AI providers and developers of autonomous agents. Claude was built to respond to human prompts, not to serve as a backend for millions of continuously running agents. Tools like Manis and OpenClaw consume far more compute power and burn through tokens at a rate that strains the underlying infrastructure.
Anthropic has already taken steps to curb demand. During peak periods, the company imposed a five‑hour session cap on its models, citing in an email to customers that “these tools put an outsized strain on our systems.” The new pricing structure is a further attempt to manage the surge in heavy‑use workloads while protecting the sustainability of its services.
The move also plays out against a backdrop of rivalry with OpenAI. Earlier this year, OpenAI hired Peter Steinberger, the creator of OpenClaw, to accelerate the rollout of AI agents to a broader audience. Steinberger responded to Anthropic’s announcement on X, criticizing the timing and suggesting the company was copying popular features into a closed system before locking out open‑source alternatives.
Analysts predict that friction between heavy‑use agent developers and AI labs will intensify. Agent tools can run for hours, interact with applications like Gmail, Slack and iMessage, and execute actions autonomously at a speed and scale far beyond human capability. That power translates into higher operational costs, prompting providers to shift more of the expense onto the end users.
Anthropic’s latest policy underscores a shifting landscape where AI companies increasingly monetize compute intensity. As developers continue to build sophisticated autonomous agents, pricing models that separate basic access from heavy usage are likely to become the norm.