On May 14, Anthropic announced that from June 15, 2026, Claude subscriptions will split into two billing pools. Usage through Anthropic’s first-party tools — the official Claude Code CLI and the chat interface — keeps drawing on the existing subscription cap. Usage through anything else — Zed, Cursor, OpenClaw, your own SDK harness, claude -p invoked from a foreign script — moves into a new “Agent SDK credit” capped at the subscription tier and billed at full API rates once exhausted.
I have been waiting for this shoe to drop since Anthropic’s quiet test in late April that removed Claude Code from the $20 Pro plan. The split is the formalization.
The Numbers
Per Zed’s own writeup of the change, the new credit pool per tier is:
| Tier | Subscription | Monthly Agent SDK credit |
|---|---|---|
| Pro | $20/mo | $20 |
| Max 5x | $100/mo | $100 |
| Max 20x | $200/mo | $200 |
Zed describes the previous arrangement as “roughly 15-30x compared to API pricing” — meaning a Max 20x subscriber running long-cycle agent workflows through Zed or a custom SDK harness was getting their token consumption subsidized by that multiplier. After June 15 the subsidy stops at the dollar value of the credit, and overage is billed at the API rate card.
Who Gets Hit
If your stack looks like this, nothing changes:
- Claude Code CLI run interactively
- claude.ai chat
- Claude Desktop app
If your stack looks like this, your bill changes:
- Claude Code routed through Zed via ACP
- Cursor’s Claude integration via the SDK
- OpenClaw or any custom Agent SDK harness
- CI/CD agents that authenticate with a subscription account via SDK
claude -pinvoked from automation scripts that consume tokens at scale
The distinguisher is the Agent Client Protocol and direct SDK usage. If the third-party tool talks to Anthropic through ACP or the SDK rather than shelling out to the official claude binary, it draws from the capped credit. The official Claude Code CLI keeps its existing subscription budget.
Why This Is Happening
Anthropic has not posted an explicit rationale on the public blog. Axios’s reporting cites posts from Anthropic’s Alex Albert and Noah Zweben on X, and frames the change against an environment where some subscribers were paying $20-$200 per month and pulling hundreds or thousands of dollars in tokens through long-running agent loops. A 15-30x subsidy was never going to survive contact with serious agentic workloads. The split caps the implicit subsidy at a known dollar figure per tier and pushes everything past that to API rates.
The competitive context, per the same Axios piece: OpenAI is courting agent builders with its own pricing moves. Anthropic is not exiting the segment, it is repricing it.
What This Means For Local AI
This is the same arc I keep ending up at. Frontier API economics are getting harder to amortize for agentic workloads, and the gap between “Opus 4.7 in a loop” and “Qwen3.6-27B in a loop” shrinks every time the API gets more honest about what those loops actually cost. A 4090 paid for once, running Qwen3.6 with MTP at 140 tok/s, has no June 15 cliff. It just runs.
The local model is still 10 points off Opus 4.7 on SWE-bench. That gap remains. But for the agentic use cases where you call the model thirty times per task, the subscription-subsidy era was making the comparison look more lopsided than it actually was. June 15 normalizes that.
What To Do
- Audit your stack: does it use the official Claude Code CLI, or Agent SDK/ACP under the hood?
- If you ship a Claude-backed tool that customers point at their personal subscription, warn them. Four weeks is enough notice if you start now.
- If you have a CI/CD pipeline authenticating as a user via SDK, move it to API-key billing before the credit runs out and the meter starts at a number you did not budget for.
- If you have been quietly arbitraging the 15-30x subsidy for production workloads, the arbitrage is closing. Plan accordingly.
Anthropic keeps the frontier. The economics of running that frontier inside someone else’s agent loop just changed.
Sources: