Flat Rate Is Over for Third-Party Agents
On April 4, 2026, Anthropic informed Claude Pro and Max subscribers that their plans would no longer cover usage from third-party agent frameworks, starting with Openclaw. The policy took effect with less than 24 hours of advance notice. The message was direct: subscription capacity is reserved for Anthropic’s own products. Third-party tools need API keys and pay-as-you-go billing.
Openclaw had become one of the fastest-growing open-source projects in the AI space after gaining traction in late 2025. Nvidia CEO Jensen Huang described it as potentially the most popular open-source project in human history. In crypto, developers adopted it quickly for autonomous workflows: monitoring wallets, executing trades, managing DeFi positions, and running onchain automations around the clock.
Boris Cherny, Head of Claude Code at Anthropic, confirmed the policy on X, framing it as a capacity decision. Third-party harnesses bypass prompt caching and other efficiency optimizations built into Anthropic’s own interfaces, putting what the company described as an “outsized strain” on its systems. Cherny also confirmed that enforcement will expand to all third-party harnesses beyond Openclaw in the coming weeks.
What Developers Can Do
Anthropic offered three mitigation options. Users can enable pay-as-you-go extra usage billing and continue running Openclaw workflows. Alternatively, they can switch to direct API keys, which Cherny noted would benefit from better prompt caching. As a one-time gesture, Anthropic offered a credit equal to one month’s subscription price, redeemable by April 17, 2026, plus discounts of up to 30% on pre-purchased extra usage bundles. A full refund option is also available.
The core subscription remains intact for Claude.ai, Claude Code, and Claude Cowork. Enterprise and Team plan treatment under the new policy has not been officially clarified.
The Cost Shift and the Community Response
The financial impact is the central issue. Flat-rate subscriptions made heavy agentic workflows viable for individual developers and small teams. Under metered billing, a single autonomous agent running continuously can generate estimated costs of $1,000 to $5,000 per day in extreme usage cases. For those who built production-grade crypto automations assuming unlimited subscription access, this is a material change in operating economics.
Openclaw’s creator, Peter Steinberger, had been in discussions with Anthropic about the rollout and reportedly delayed enforcement by roughly a week through those talks. Steinberger joined OpenAI in February 2026, which drew speculation online. Cherny addressed this directly, stating the policy is driven by engineering constraints rather than competitive positioning. He also submitted pull requests to improve Openclaw’s cache hit rates for developers migrating to API access.
“Third-party harnesses bypass prompt caching and other efficiency optimizations, putting an outsized strain on systems we manage carefully.” — Boris Cherny, Head of Claude Code, Anthropic
Community reaction split along predictable lines. Developers who built heavy automations on top of subscriptions called it a bait-and-switch. Others accepted the logic: a 24/7 autonomous trading agent consuming tokens nonstop is a fundamentally different product than a human having a conversation. Some users are already migrating to competing model providers or local model setups to keep costs flat.
Broader Context
Anthropic has been expanding agentic capabilities inside its own products, including loop functions and scheduled task features in Claude Code and Claude Cowork. These features parallel what Openclaw provides externally. The company has not published a dedicated policy page or updated its terms as of April 5. Communication happened through direct subscriber email and Cherny’s X posts, a minimal rollout for a change with significant downstream impact on the developer community that built most aggressively on top of Claude.
