Banning OpenClaw, launching their own Agent platform: Anthropic's infrastructure ambitions come to light

Author: Claude, Deep Tide TechFlow

Deep Tide Quick Read: Anthropic makes two moves in parallel this week: on April 4, it cuts off access to subscription quota for 135k OpenClaw instances, and on April 8, it launches a cloud Agent hosting service called Managed Agents.

When these two steps are combined, you can see a clear shift from selling model APIs to selling the infrastructure for running Agents. With Anthropic having just crossed $30 billion in ARR, it is redefining the rules of the AI Agent game using pricing power and platform lock-in effects.

Anthropic completed two actions within a week. Put together, the intent is so clear that it hardly needs interpretation.

On April 4, Anthropic officially cut off Claude Pro and Max subscribers’ ability to use quota via third-party Agent frameworks such as OpenClaw. The 135k active instances were forced overnight to move to pay-as-you-go or API-based billing. Four days later, on April 8, Anthropic released the Claude Managed Agents public beta, providing the full suite of cloud-hosted infrastructure, from sandbox execution to state management and multi-Agent coordination.

Shut the door on one side, open the shop on the other. The open-source community’s anger is understandable, but from a business-logic standpoint, these two moves serve the same goal: Anthropic no longer wants to be only a model supplier—it wants to be an infrastructure platform for the Agent era.

Banning OpenClaw, and ending the $20 “all-you-can-eat” subscription

OpenClaw’s hype needs no elaboration.

Previously, some users ran Agents using the $20 monthly subscription quota for Claude, but the problem was the economics. The compute cost consumed per day by a single heavy user can reach $1,000 to $5,000, which obviously increases the burden on Anthropic.

According to a report by VentureBeat, Boris Cherny, head of Anthropic Claude Code, announced this change on the X platform, saying the subscription plan “was never designed for usage patterns of third-party tools,” and that the company needs to “prioritize serving customers who use our own products and APIs.”

The timeline makes the matter more nuanced.

Earlier this year, in January, Anthropic raised objections to Clawdbot’s trademark. On February 14, Steinberger announced he was joining OpenAI, and Sam Altman publicly welcomed him. On February 20, Anthropic updated its service terms, explicitly prohibiting the use of subscription OAuth tokens for third-party tools. On April 3, Semafor reported that Anthropic is building its own OpenClaw competitor; Chief Business Officer Paul Smith admitted that customers have “always been asking us to do this.” April 4—full execution.

Steinberger’s response was quite direct: “First copy the hot open-source features into your own closed tool, then lock the open source out the door.” He and investor Dave Morin tried to negotiate with Anthropic; at most, they managed to get the rollout delayed by a week.

Anthropic offered two transition measures: a one-time grant of quota equivalent to the price of a monthly subscription, and pre-purchasing additional usage packs with up to a 30% discount. But for heavy users, switching from a fixed monthly fee to pay-as-you-go pricing could cause costs to skyrocket by 50x.

Managed Agents: From selling models to selling runtime

In the same week that it moved to ban OpenClaw, Anthropic provided its own alternative.

On April 8, Claude Managed Agents entered public beta. According to Anthropic’s engineering blog, the design philosophy for this service borrows from operating system abstraction principles: break an Agent into three independently replaceable components—session (session logs), harness (call loop), and sandbox (code execution environment). The three are decoupled, so if any one crashes, it doesn’t affect the rest.

The engineering blog explains in detail why this architecture is needed. In the early versions, all components were placed in the same container, and the container turned into a “pet.” Once it crashed, the entire session was lost, and debugging couldn’t access user data.

After decoupling, the container becomes “cattle”: when it crashes, you replace it with a new one. The harness restores state from the session logs and keeps running.

In terms of pricing structure, Managed Agents charges an additional $0.08 per session runtime hour on top of standard API token fees (billed per millisecond). Idle waiting time is not billed. Web searches triggered by the Agent are charged at $10 per 1,000 searches.

According to a report by SiliconANGLE, companies including Notion, Rakuten, Asana, and Sentry have become early users. Asana embeds the Agent into its project management workflow, building “AI teammates” that can automatically claim tasks and draft deliverables. Sentry pairs its existing debugging Agents with patch-generation Agents powered by Claude; the process from discovering bugs to submitting pull requests was shortened from the originally planned several months to a few weeks.

Currently, two features are in research preview: first, the Agent can spin up sub-Agents when handling complex tasks; second, the Agent’s self-evaluation capability—after developers define what “success” means, Claude iterates on its own until it meets the criteria.

The platform economics behind the two moves

Put the two things side by side, and the business logic is very clear.

Anthropic’s ARR has just crossed $30 billion. According to The Information, this figure is more than double the roughly $9 billion at the end of 2025, and it exceeds $1 million in annual spend for over 1,000 enterprise customers.

Claude Code alone contributes over $2.5 billion in annualized revenue. At this scale, it’s not sustainable to have 135k OpenClaw instances consume compute worth thousands of dollars with a $20 monthly fee.

But cost control alone isn’t enough to explain why Managed Agents was launched at this time.

In an interview, Angela Jiang, head of platform products at Anthropic, said there is still a gap between the capability of the company’s models and how enterprises actually use them. The goal of Managed Agents is to enable enterprises to deploy “teams of Claude Agents” to handle real work workloads.

This is a classic platform lock-in strategy. Once an enterprise’s Agents run on Anthropic’s hosted infrastructure, data pipelines, monitoring configurations, and the permission system are embedded into day-to-day workflows, and migration costs rise sharply.

For a company valued at $380 billion and considering an IPO, this kind of stickiness is far more valuable than simple API call fees.

Previously, multiple analysts and social media influencers have expressed the view that “the true battlefield of AI is in orchestration.” Whoever controls an Agent’s routing and combination logic makes the underlying model substitutable.

OpenClaw supports switching among multiple models such as Claude, GPT-4o, and Gemini. After 135k users are forced to leave fixed-rate pricing, some people will shift to local models or other providers.

In February of this year, Google took a similar action by banning third-party tools from borrowing OAuth authentication from the Gemini CLI. Put these together, and you can see that the AI industry is moving from “model competition” to “platform competition.”

The subscription-style unlimited model is coming to an end across the industry; pay-as-you-go billing and infrastructure bundling will become the new norm.