Anthropic announced on Thursday a significant change in its subscription model. It is restoring access for third-party autonomous artificial intelligence (AI) agents while simultaneously ending the era of “compute arbitrage” that previously allowed power users to exploit flat-rate pricing.
The company revealed on X that it is introducing a new Agent SDK credit system. This move effectively reverses an April policy that prohibited Claude Pro and Max subscribers from using their accounts to power external, non-Anthropic harnesses like OpenClaw.
The initial ban followed a period of unsustainable infrastructure strain. According to Boris Cherny, head of Claude Code, third-party tools were often unoptimized, bypassing Anthropic’s “prompt cache” mechanisms designed to save compute cycles. This led to a financial mismatch, in which users paying $20 to $200 monthly consumed thousands of dollars in token value through inefficient autonomous workflows.
Despite Anthropic’s expansion into the 300MW Colossus 1 data center housing over 220,000 GPUs, the demand for agentic workflows had begun to outpace the supply of sustainable compute.
The restoration of access comes with a strict technical catch, however. Programmatic usage is no longer drawn from the general subscription pool. Instead, users receive a fixed monthly credit — billed at standard API rates — commensurate with their plan. For example, Pro Plans cost $20 in Agent SDK credits; Max 5X Plans are $100 in Agent SDK credits.
These credits follow a “use it or lose it” model and do not roll over. Once exhausted, programmatic tasks – such as running GitHub Actions or third-party tools via the claude -p command – will cease unless the user enables pay-as-you-go extra usage billing. Interactive tasks, like chatting in a browser or using the terminal-based Claude Code, remain covered under the standard high-capacity limits.
Anthropic framed the shift as a “simplification” that legitimizes third-party ecosystems, but the developer community has reacted with sharp criticism.
“AI vendors are formally separating interactive consumption from autonomous agent consumption, with distinct metering for each. The split reflects frontier labs converging on a common position: always-on agents are infrastructure workloads, and subscription economics will not subsidize unbounded programmatic agents,” said Mitch Ashley, vice president and practice lead for Software Lifecycle Engineering at The Futurum Group. “Enterprises building on agent SDKs now treat agent workloads as a budgeted infrastructure line with cost attribution per workflow. Inefficient code and unbounded autonomy surface directly as spend, putting FinOps controls and AgentOps observability on the procurement checklist alongside model performance.”
Prominent AI developer Theo Browne warned that for users of external tools like Zed or Conductor, the change represents an effective usage cut of “25x.” Similarly, former Meta Platforms Inc. engineer Kun Chen described the move as “pulling the plug” on programmatic value, suggesting it may drive elite talent toward competitors like OpenAI.
“They’re disguising this as ‘free credits’. Don’t fall for it,” Browne cautioned his followers.
By aligning subscription programmatic usage with its Developer Platform (API) pricing, Anthropic is protecting its margins and ensuring system stability. While the move ends the “all-you-can-eat” model for autonomous agents, it provides a stable, metered sandbox for developers who were previously locked out entirely.
For Anthropic, the priority is clear: managing the physical limits of compute in an increasingly agent-heavy world, even at the cost of power-user goodwill.
