Anthropic Splits Claude Agent SDK Into Separate Credit Pool

Anthropic announced changes that will fundamentally restructure how Claude subscribers can use the Agent SDK and third-party AI agent tools. Starting June 15, 2026, programmatic Claude usage — including the Agent SDK, the claude -p CLI command, and third-party apps built on the SDK — will draw from a separate, dedicated monthly credit pool instead of the general subscription quota that interactive users get.

The change ends a months-long policy fight that started when Anthropic banned third-party agents from subscription plans entirely in April, and it lands the same week OpenAI and xAI are aggressively courting developers with new agent tooling.

The New Numbers

Each subscription tier gets a fixed monthly credit allocation specifically for programmatic use:

The Pro plan gets $20 in monthly Agent SDK credit. The Max 5x plan gets $100. The Max 20x plan gets $200. Team plans get $100 per seat per month. Enterprise plans get $200 per seat per month. Credits are metered at standard API list rates — the same rates a developer would pay calling the Claude API directly.

Critically, these credits do not roll over. Unused credit at month-end disappears. And if a developer burns through their monthly allocation, they can't fall back on their general subscription pool to keep going. They have to buy additional usage credits at API rates, separately.

That asymmetry matters. Heavy automation users effectively now have two separate budgets — one for chatting with Claude in the web app, and a smaller, harder-capped one for running agents.

How We Got Here

The path to this announcement has been bumpy. In April 2026, Anthropic banned third-party agent tools from drawing on Claude subscriptions, citing the high computational costs of long-running autonomous workflows. That move broke a lot of indie developer setups overnight and triggered backlash in the developer community.

Anthropic briefly tested removing Claude Code itself from the Pro plan, which generated even sharper pushback. The company then walked back the broadest restrictions and started shaping a compromise — which is what landed today.

The economic pressure behind the change is real. Agent workloads are different from interactive chat. A developer using Claude in the web app might burn through a few thousand tokens per question. An agent running for an hour can chew through millions, often calling the model recursively as it plans, executes, and revises its work. Pricing those workloads under a fixed-price subscription stops working as soon as a meaningful share of users start running agents seriously.

Why the Old Model Stopped Working

Under the previous unlimited-within-the-plan structure, the math only worked because most subscribers weren't running heavy agents. The Pro plan's value proposition assumed a typical interactive user — not someone with five Claude Code sessions running in the background overnight.

By splitting agent usage into a separate metered pool, Anthropic protects itself against heavy automation users whose actual cost dwarfs their subscription fee, while keeping interactive subscription pricing where it is for the much larger group that never touches the SDK.

What This Means for Different Users

For casual SDK users — someone running an occasional automation, prototyping a small agent — the $20 Pro credit will cover meaningful work. Twenty dollars at API rates is enough for thousands of small agent runs or a handful of complex multi-step workflows.

For production users — developers running agents inside scripts, businesses with always-on automations, anyone using claude -p in CI pipelines — the math changes sharply. A single long agent run can easily consume several dollars of API credit. Running multiple such jobs daily blows through even the $200 Max 20x credit quickly. These users will likely move to direct API billing, where the only meaningful cap is what they choose to spend.

For third-party agent tool builders, the change is mixed. Their tools can now legitimately run on Claude subscriptions again — that's an improvement on the April ban. But the cap structure means their users may hit limits faster than expected, which translates to support tickets and user frustration that the tool builders will have to absorb.

The Competitive Picture

OpenAI is moving in a different direction. The company is currently pricing aggressively on agent and Codex usage, including the rollout of Codex on the ChatGPT mobile app available even on the free tier. Where Anthropic is tightening, OpenAI is loosening — at least for now.

xAI's new Grok Build CLI is gated behind the $300/month SuperGrok Heavy tier, which is a different bet entirely: lower volume of users, each paying enough to cover heavy automation usage by default.

The three approaches frame the same question — how to price autonomous agent workloads sustainably — and Anthropic's split-pool model is the most explicit attempt to separate the economics of agent usage from interactive chat usage.

What Developers Are Saying

Reactions have been mixed. Some developers welcome the clarity, especially after the April ban left third-party tool integrations in limbo. Others note that the credit caps work out to less compute than they were effectively using under the old structure. The non-rollover provision has drawn particular criticism — agent usage isn't uniform across a month, and resetting the allocation leaves credit on the table during quiet stretches that could have funded peak weeks.

What's Next

The June 15 implementation date gives developers a month to reorganize their workflows. The most likely shifts: heavier production users will move to direct API billing, indie builders running modest automations will stay on subscriptions and adapt to the cap, and third-party agent tools will publish guidance on how to keep usage within the new limits.

Anthropic is also expected to clarify how the credit interacts with promotional credit, free tier access, and overage billing — the documentation is still light on edge cases.

The Bottom Line

This is Anthropic putting a real price on what was effectively a subsidy. Under the old structure, heavy agent users were getting compute that cost Anthropic far more than they paid. The new credit pool ends that. For the developer community, the change is going to feel like a price increase even though sticker subscription prices haven't moved — because the value of the subscription, for the people who used it most aggressively, has been quietly cut in half. Whether that's the right call depends on how much margin Anthropic was leaving on the table to keep agent workflows artificially cheap. The June 15 transition will tell us how many heavy users decide the new math doesn't work for them.