Anthropic announced a policy reversal on its official developer account @ClaudeDevs: paid subscribers now get a dedicated "Agent SDK credits" allowance for third-party autonomous agent tools such as OpenClaw. The move looks like a concession to developers, but it effectively redraws the financial boundary—programmatic call costs are no longer borne by the subscription fee pool.
Why the Block, and Why the Unblock
In early April, Anthropic banned subscribers from using Claude with OpenClaw and other non-official agent frameworks, citing "capacity and service stability." The real reason was financial: some users paid only $20 to $200 in fixed monthly fees but consumed hundreds or thousands of dollars worth of tokens through autonomous agents. Boris Cherny, head of Claude Code, explained that Anthropic's own tools are optimized for prompt caching, reusing processed text to lower compute costs. Third-party tools like OpenClaw bypass this cache, forcing near-fresh computation per turn—unsustainable for Anthropic. Even with over 220,000 GPUs at the Colossus 1 data center, agent workload demand outgrew infrastructure capacity.
How the New Agent SDK Credits Work
Each subscription tier gets a monthly credit cap: Pro $20/month, Max 5x $100/month, Max 20x $200/month, Team (Premium) $100/seat/month, Enterprise (Premium) $200/seat/month. When users run claude -p, trigger a GitHub Action, or connect to tools like OpenClaw, the system switches to the Agent SDK credit pool for billing at standard API rates (per million tokens). Once credits are exhausted, programmatic access pauses; unused credits do not roll over. Anthropic engineer Lydia Hallie added operational details on X.
Compute Arbitrage Window Closes; Community Reacts
The immediate impact is the end of the "compute arbitrage" phenomenon that emerged in early 2026—using a $20 Pro subscription to run autonomous agent workflows that would cost hundreds of dollars via the standard API. Theo Browne of T3.gg warned developers: "If you use OpenClaw with Claude subscription, your usage might be cut by 25x." Atlassian senior engineer Kun Chen called it "Anthropic officially pulling the plug on all programmatic subscription use." Ben Hylak, CTO of Raindrop.ai, was more blunt: "Either this is very stupid, or it shows how badly Anthropic is hurting for GPU resources."
By explicitly allowing Conductor, OpenClaw and similar tools through Agent SDK authentication, Anthropic endorses the third-party ecosystem. But the independent credit pool design shifts the marginal cost of inefficient usage back to end users instead of the common subscription pool. Expect many competitors to follow the same playbook—compute remains scarce.