Two months ago, Anthropic pulled a quiet but disruptive policy move, cutting off third-party agents like OpenClaw from its Claude subscription plans without warning. The ban fractured developer communities that had built workflows around external Claude clients. On Tuesday, Anthropic reversed course. Starting June 15, third-party agents will be permitted again, but under a new structure: a monthly "Agent SDK credit" pool separate from standard subscription usage, sized by plan tier.
How the Ban Happened
Anthropic's April decision to disable third-party harnesses came after internal analysis showed external agents were extracting far more compute value than their subscription price reflected. Community analysis documented OpenClaw users consuming roughly $236 of API-equivalent compute per month from a $20 Pro subscription, a ratio of twelve to one. At the high end of Max 20x usage, independent estimates put the implicit subsidy at 175 to one. The result was degraded service for all subscribers during peak hours, and Anthropic concluded the flat-rate system could not sustain the load.
The ban generated immediate backlash. Developers who had built continuous-integration pipelines and autonomous coding loops around OpenClaw and similar harnesses found their tools broken overnight. Several high-profile Hacker News threads catalogued the damage. Anthropic offered no clear timeline for reinstatement, and by May many affected users had migrated to direct API access or competitor models.
Agent SDK Credit Pool at a Glance
- Ban dateApril 2026
- Reinstatement dateJune 15, 2026
- Pro plan credit$20/month
- Max 5x credit$100/month
- Max 20x credit$200/month
- Credit rolloverNone — expires monthly
- Overflow billingOpt-in only
How the Credit Pool Works
Under the new system, interactive Claude Code in the terminal remains part of the standard subscription allowance and is unaffected. What changes on June 15 is programmatic access: Agent SDK calls, claude -p, automated GitHub Actions runs, and third-party harnesses will draw from the dedicated credit pool at standard API list rates. When the pool runs out, automated requests halt. There is no automatic overflow to the subscription limit and credits do not carry to the next month.
Users who need more than the monthly credit can enable overflow billing through account settings. Anthropic designed this as opt-in, citing the lesson of earlier this year when a customer accidentally burned through $500 million in Claude usage without realizing the exposure. The credit pool architecture makes the cost model explicit before usage begins, not after.
"Starting June 15, 2026, programmatic usage of Claude via subscription plans...will draw from a separate monthly Agent SDK credit instead." Anthropic, @ClaudeDevs, June 2026
The Larger Picture
This is the third significant adjustment to Anthropic's subscription terms in 2026. In February, the company tightened limits on Claude Code usage after power users discovered ways to run near-continuous sessions. The April agent ban followed. Now the credit pool threads a compromise: access is restored, but the cost of inefficient third-party code falls on the user rather than being absorbed into Anthropic's margins ahead of its planned IPO.
For the third-party Claude ecosystem, the announcement signals survival with tighter boundaries. OpenClaw, whose user base had scattered in recent weeks, confirmed post-announcement that it would support the new credit system. Whether $20 per month in Pro is enough to support meaningfully agentic workflows is a separate question. For light automation, it will be. For teams running pipelines that call Claude hundreds of times daily, the math requires a closer look at what the June 15 billing change actually costs them.
