Policy Shift: Reintroducing Programmatic Agent Use
Anthropic communicated the update through its official developer channels on X (@ClaudeDevs) on May 13, 2026. The move revokes a restriction imposed in early April 2026 that had prohibited using paid Claude subscriptions to run non-Anthropic agents and harnesses, citing concerns over capacity and service stability.
The initial prohibition stemmed from an unsustainable resource drain. Subscribers were reportedly paying monthly fees ranging from $20 to $200 under Anthropic’s Claude Pro and Max plans, but these accounts were simultaneously consuming token amounts potentially reaching thousands of dollars through external agents like OpenClaw. This consumption pattern created a financial and compute challenge for Anthropic’s limited infrastructure.
While the company previously redirected users to pay via its API (billed per million tokens) or purchase extra usage credits, the new system integrates third-party agent capabilities directly into the subscription structure—but with strict conditions.
The Mechanics of Agent SDK Credits
The restoration of access comes with a significant caveat: programmatic tasks are no longer subsidized by the general subscription pool. Instead, they are restricted to dedicated, fixed monthly “Agent SDK” credits, which are billed at standard API rates and range from $20 to $200 depending on the plan.
Crucially, these credits do not roll over; any unused amount expires at the end of the month. Furthermore, once a user exhausts their dedicated Agent SDK credit limit, they cannot dip into their general subscription usage limits for additional programmatic tasks and must purchase extra usage credits.
Technical Constraints Behind the Policy
To understand this structural change, experts point to Anthropic’s internal engineering requirements. First-party tools developed by Anthropic, such as Claude Code and Claude Cowork, are optimized to maximize “prompt cache hit rates”—a method that reuses previously processed text to conserve expensive compute cycles. Third-party services like OpenClaw, which run agents through external platforms (such as Discord or Telegram), were often inefficient and bypassed these crucial internal caching mechanisms.
“Boris Cherny, Head of Claude Code,” noted that the sheer volume of data being reprocessed by unoptimized third-party tools was unsustainable for Anthropic’s system stability. Even with access to massive hardware like the 300MW Colossus 1 data center and its 220,000+ GPUs, the demand from agentic workflows was outpacing sustainable supply.
The new credit system resolves this technical bottleneck by making users directly accountable for inefficiency. By allocating a dedicated dollar-amount credit, Anthropic ensures that inefficient third-party code drains the user’s specific budget rather than exceeding the value of the company’s fixed monthly subscription tiers.
Dedicated Credit Allocation Table
Anthropic has segmented access into a new hierarchy of programmatic power across its billing tiers. These dedicated Agent SDK credits are supplementary to standard usage within Anthropic products (like Claude Code).
- Pro: Dedicated credit of $20, intended for individual scripts and light SDK use.
- Max 5x: Dedicated credit of $100, suited for moderate agentic automation.
- Max 20x: Dedicated credit of $200, designed for professional-grade development environments.
- Team (Premium): Dedicated credit of $100 per seat, intended for collaborative team automation.
- Enterprise (Premium): Dedicated credit of $200 per seat, reserved for high-scale enterprise usage.
