The billing split is live. As of June 15, 2026, programmatic Claude usage — the Agent SDK, claude -p, Claude Code GitHub Actions, and every third-party tool authenticating through the Agent SDK — no longer draws from your subscription’s shared compute pool. It now runs against a separate monthly credit bucket at full API list prices.
We covered the mechanics of the change when Anthropic announced it in May. What we could not cover then is the subsidy math that made the change inevitable, the sequence of events that led here, and the competitive reaction it triggered. Now that the numbers are public and the change is live, it is worth understanding all three.
The Subsidy Was Enormous#
The simplest way to understand why Anthropic had to act is the ratio.
Independent community analyses found that OpenClaw users on a $20 Pro subscription were consuming roughly $236 of API-equivalent compute per month — a 12-to-1 ratio. For Max 20x subscribers running Opus-class models on heavy automated workloads, estimates put the effective subsidy at 175-to-1: $200/month in subscription fees against compute costs that could exceed $35,000 at API list prices.
These were not edge cases. At the time Anthropic banned third-party OAuth access on April 4, more than 135,000 OpenClaw instances were running. Multiplied across any non-trivial usage pattern, the aggregate subsidy was material.
Boris Cherny, Head of Claude Code at Anthropic, gave the official explanation: third-party agents bypass the prompt-cache optimizations that Anthropic’s own tools use. Claude Code in interactive mode is designed to aggressively cache repeated context — system prompts, CLAUDE.md contents, file structures that appear across multiple turns. External tools making direct API calls process full context from scratch on every call. The infrastructure economics are meaningfully different.
That explanation is credible. It is also incomplete. The simpler version is that Anthropic was running a $200/month service that cost orders of magnitude more to deliver to its heaviest programmatic users, and the subscription model never priced for that.
The Three-Strike Timeline#
The June 15 billing split is the fourth intervention in a six-month sequence:
January 2026: Anthropic blocked subscription OAuth tokens from being used in third-party tools. Developer backlash was immediate and sharp. The change was reversed within days.
February 2026: Terms of service revised to restrict OAuth token use to Claude Code and Claude.ai only. Enforcement was inconsistent. The ecosystem largely continued operating.
April 4, 2026: Anthropic banned third-party agents from using subscription credentials outright. OpenClaw, Conductor, Zed’s ACP integration, and Jean were all caught in the action. Theo Browne, whose Conductor tool was among those blocked, calculated a 25x effective usage cut for his users and publicly canceled his subscription. He called the ban “an attack on open-source tooling that repudiates months of explicit promises from Anthropic’s developer relations team.”
June 15, 2026: The credit pool separation goes live, replacing the ban with a metered structure. Third-party tools can resume using the Agent SDK, but now against a finite dollar-denominated credit that does not scale with subscription tier beyond the plan credit amount.
The sequencing reflects a company that moved incrementally, backed off when pushback was severe, and eventually landed on a structure it can defend. The April ban was untenable because it punished legitimate tool developers with no path to compliance. The June billing split gives those developers a path — pay API rates via the credit pool — while closing the unlimited subsidy.
Sam Altman’s Timing Was Not an Accident#
The announcement that Anthropic would implement the billing split was made on May 14, 2026. On that same day, OpenAI CEO Sam Altman publicly offered enterprise customers two months of free Codex usage if they switched from Claude Code within 30 days.
The offer was targeted. The framing was explicit. Altman was not announcing a new product feature — he was naming Claude Code by category and offering the subsidy that Anthropic was withdrawing.
This is the market dynamic that makes the billing split more complicated than a straightforward infrastructure economics story. Anthropic is withdrawing a subsidy that, at scale, made Claude Code effectively the cheapest programmatic AI coding infrastructure available. OpenAI, which does not have equivalent per-seat subscription exposure, can temporarily absorb that cost as a customer acquisition expense.
The competitive window is narrow but real. For enterprise customers evaluating whether to migrate CI pipelines or automated agent workflows, the math in June 2026 briefly favors Codex in a way it did not in April.
What the Credits Actually Buy#
The credit amounts are:
| Plan | Monthly subscription | API credit |
|---|---|---|
| Pro | $20 | $20 |
| Max 5x | $100 | $100 |
| Max 20x | $200 | $200 |
At Opus 4.8 pricing ($5 input / $25 output per million tokens), a $200 Max 20x credit covers approximately 40 million input tokens or 8 million output tokens. A demanding agent session — loading codebase context, running multi-step reasoning, generating and testing code — commonly burns 500,000 to 1,000,000 tokens. The $200 credit funds roughly 8 to 40 sessions at Opus rates per month before overflow billing at API list prices begins.
For lighter workloads using Haiku 4.5 ($1 input / $5 output), the same $200 funds roughly 200 million input tokens — the arithmetic changes materially.
There is also a critical operational note: credits must be claimed. They do not activate automatically on June 15. Users who did not navigate to their account settings and claim credits before today are not covered. This detail generated its own wave of developer complaints when the requirement became clear.
The Community Notes Correction#
When Anthropic posted about the billing split on X, framing it as a way to give developers “more programmatic compute,” the response included a Community Notes correction added within hours. The note clarified that the change represented a reduction in accessible compute for most programmatic use cases relative to what had been available under the subscription model.
Community Notes corrections on company announcements are relatively rare. They require community consensus that a post is misleading. The correction landing on Anthropic’s own announcement post underscores how the change landed with the developer community that reads the subsidy numbers honestly.
The Honest Assessment#
Anthropic had to make this change. The 175-to-1 subsidy is not a sustainable business model, and the alternative — continuing to fund unlimited programmatic compute via flat-rate subscriptions — would eventually have required either much higher subscription prices or capacity constraints that degraded the interactive experience.
The execution was imperfect. The January reversal, the April ban, the unclear communications, the require-to-claim credit mechanics — these created unnecessary friction with a community that Anthropic needs to remain onside. The Developer Trust article that circulated after the April ban reflected a genuine erosion that the billing split, even with a clear economic rationale, does not fully repair.
What the billing split does accomplish is establish a durable structure. Interactive Claude Code remains on the subscription. Programmatic usage is metered. The economics are transparent. Developers building on the Agent SDK now know what they are building on: API pricing, not a subsidized subscription pool.
For most professional use cases, the credit amount covers what you need. For teams running high-volume automated agent workflows, the math requires revisiting budget allocations — and probably a conversation about whether Haiku 4.5 is adequate for tasks currently running on Opus.
The subsidy era is over. The infrastructure era is beginning.
Credit pool amounts and plan breakdowns confirmed via Codersera and The New Stack. The OpenClaw 135,000-instance figure and Theo Browne quotes are from public developer commentary following the April 4 ban. Overflow billing applies at standard API list prices once the monthly credit is exhausted.