Anthropic is carving Claude into two billing lanes. Starting June 15, Agent SDK and claude -p usage will no longer draw from normal Claude subscription limits. Each plan tier gets a fixed monthly Agent SDK credit instead: $20 for Pro, $100 for Max 5x, $200 for Max 20x and Enterprise Premium seats. Once the credit is gone, additional usage bills at standard API rates, but only if the user has opted in to usage credits. Otherwise, requests stop until the next billing cycle.

This follows April’s cut-off, when Anthropic blocked Claude Pro and Max subscriptions from third-party agent harnesses like OpenClaw, OpenCode, and Pi. Boris Cherny, head of Claude Code, attributed the restriction to engineering constraints. “Our systems are highly optimized for one kind of workload, and to serve as many people as possible with the most intelligent models, we are continuing to optimize that,” he said on X at the time.

Two Moves, One Pattern

The April restriction drew a line between Anthropic’s own surfaces (Claude.ai, Claude Code, Cowork) and everything else. The June billing change extends that logic into Anthropic’s own agent tooling: even the Agent SDK, which Anthropic builds and maintains, now gets metered separately from interactive Claude usage.

The combined effect is a three-tier Claude access model:

  • Tier 1 (subscription-covered): Interactive Claude Code, Claude.ai, Cowork. Full subscription limits apply.
  • Tier 2 (capped credit): Agent SDK, claude -p, Claude Code GitHub Actions, third-party Agent SDK apps. Draws from a fixed monthly pool, then API rates.
  • Tier 3 (API-only): Third-party harnesses like OpenClaw that don’t use the Agent SDK. Subscription access revoked entirely in April. API key required.

For a Pro subscriber paying $20/month, the Agent SDK credit is exactly $20. That means a developer running moderate agentic workloads could exhaust the credit in a few days of heavy claude -p usage, then face API rates for the rest of the month. The credit structure is sized for “individual experimentation and automation,” per Anthropic’s help center. Production workloads are explicitly directed to API billing.

The IPO Lens

Independent analyst Vincent Schmalbach argues this pattern marks Anthropic’s transition from AI lab to public-market-ready company. The moves are consistent with what a company does before an IPO: protect margins, create predictable revenue streams, and segment products to maximize average revenue per user.

The flat-rate subscription model was always a loss leader for heavy agent workloads. A developer running 100 claude -p calls a day generates far more inference cost than someone chatting interactively. By splitting billing, Anthropic can report subscription revenue and metered API revenue as distinct, growing line items, exactly the kind of metric public market analysts reward.

Schmalbach also points to Anthropic’s commercial terms, which prohibit customers from using the service to build competing products, train competing models, or resell services without approval. Combined with the 2025 incident where Anthropic revoked OpenAI’s API access and restricted Windsurf’s direct access, the picture is of a company that will enforce platform boundaries aggressively.

What Changes for Agent Builders

Every agent framework that relies on Claude now faces a cost recalculation. OpenClaw users already lost subscription access in April. Developers using claude -p in CI pipelines, shell scripts, and automation workflows will hit the credit cap and shift to metered billing after June 15.

The practical question is whether the Agent SDK credit covers real workloads. For a solo developer running occasional agentic tasks, $20 might be enough. For a team running shared production automation, Anthropic’s own documentation says the answer is no: “Teams running shared production automation should use Claude Platform with an API key for predictable pay-as-you-go billing.”

The broader signal is that Claude as a flexible, general-purpose developer tool is becoming Claude as a segmented product line. Interactive use stays cheap and subscription-covered. Agentic use gets capped, then metered. Third-party harness use gets API-only billing. Each step pushes developers toward either Anthropic’s own surfaces or higher spend.

For builders choosing their LLM stack, the calculus now includes not just capability and price per token, but how the provider segments access. Anthropic is optimizing for margin. That is a rational business decision. It is also one that every agent framework needs to price into their architecture.