Anthropic has reopened the door to third-party Claude agents on paid subscriptions, but it is not handing out an all-you-can-eat buffet. Paid users can again point their Claude plans at external agentic systems such as OpenClaw, yet programmatic usage now draws from a separate monthly pool of ”Agent SDK” credits that expires if unused.
That is the real story here: Anthropic is restoring flexibility while killing the subsidy loophole that made a flat-rate subscription behave like a much more expensive API plan. For developers, that is better than the blanket ban from early April; for Anthropic, it is a way to stop powerful automation from chewing through shared compute capacity.
How Anthropic is splitting Claude usage
The new system separates chatty, human-facing work from machine-driven automation. If you are using Claude in a browser or working interactively in Claude Code, you keep drawing from the normal subscription limits. Switch to non-interactive tasks such as the claude -p command, a GitHub Action, or a third-party tool like OpenClaw, and the Agent SDK credit kicks in instead.
- Pro: $20 monthly Agent SDK credit
- Max 5x: $100 monthly Agent SDK credit
- Max 20x: $200 monthly Agent SDK credit
- Team (Premium): $100 / seat
- Enterprise (Premium): $200 / seat
Once that credit runs out, programmatic usage stops unless the subscriber opts into extra usage billed at standard API rates. The sting in the tail is simple: unused credits do not roll over, so this is a monthly budget, not a stored asset.
Why Anthropic changed course
The original ban was less about ideology and more about math. Anthropic said third-party agents were hammering its systems in ways that first-party tools are built to avoid, especially around prompt caching, which helps reuse prior work and reduce expensive reprocessing. In plain English: some users were paying subscription prices while consuming far more model capacity than that price was meant to cover.
That problem was always going to collide with the economics of flat-rate AI. As rivals such as OpenAI and Google push harder into developer tooling, subscription bundles are becoming a battleground between convenience and infrastructure costs. Anthropic now appears to be choosing a middle lane: keep developers inside the tent, but charge them in a way that maps more closely to actual usage.
Developer backlash is predictable
Not everyone is buying the company’s ”simplification” pitch. Many power users see the update as a disguised downgrade because it replaces open-ended experimentation with a hard monthly ceiling. That is especially painful for builders who treated Claude subscriptions as a cheap front door to high-volume agent workflows.
The more charitable read is that Anthropic has accepted a harsh reality: if an agent is inefficient, the cost has to land somewhere. Under the new setup, it lands on the subscriber instead of being absorbed by the general pool. That may annoy enthusiasts, but it also gives Anthropic a better shot at protecting margins while keeping third-party agents technically supported.
What Claude subscribers can expect next
For individual developers, this is still better than a full lockout. For teams building production automation, it is a reminder that subscription plans are not meant to be silent stand-ins for enterprise API usage. The likely result is a cleaner split: Claude subscriptions for light experimentation, API billing for serious scale.
The open question is whether Anthropic can hold onto the developers who liked the old loophole without training them to look elsewhere. If the new credit buckets feel too small, the company may get the financial discipline it wants and a smaller fan club than it had before.