If your team runs Claude Code, the way you pay for it changes on 15 June 2026. Anthropic is splitting subscription billing into two buckets. Interactive use stays on your existing plan limits, while automated and agentic use moves to a separate monthly credit. The date matters because the new meter starts the moment the change lands, not when you next review your bill. For most Australian teams the practical question is simple: which of your Claude workloads are quietly running in the background, and what will they cost once they sit on their own meter?
What actually changes on 15 June
The change draws a line between two kinds of usage. Interactive work, meaning Claude chat on claude.ai and Claude Code driven by hand in your terminal or IDE, stays covered by your current Pro or Max plan. Anything automated moves to a new monthly Agent SDK credit. That second bucket is where the cost surprises live.
Interactive use stays on your plan: claude.ai chat and Claude Code used live in your terminal or editor keep drawing from existing Pro and Max limits, with no change.
Automated use moves to a credit: Claude Agent SDK calls, claude -p, Claude Code GitHub Actions, and third-party editors such as Cursor and Zed that authenticate through your subscription now draw from a separate monthly Agent SDK credit.
Credit amounts: Pro gets US$20 a month, Max 5x gets US$100, and Max 20x gets US$200 (roughly $30, $150 and $300 AUD).
When the credit runs out: usage falls back to standard API rates if you have enabled overflow, otherwise automated requests stop until the next cycle.
No rollover: unused credit does not carry into the following month.
Why Anthropic is making the split
Subscriptions were effectively subsidising agentic usage. Community estimates put the gap at fifteen to thirty times the equivalent direct API price. A chat session uses a modest amount of compute. A long-running agent loop driving an editor or a nightly job can consume far more, and that traffic grew quickly as more tools started authenticating through Claude subscriptions. The split prices automated work closer to what it actually costs to run, and keeps interactive plans predictable for the people using Claude by hand.
How to tell which bucket a job lands in
The simple test is whether a human is waiting on the output. If you are sitting in the terminal or the IDE watching Claude work, that is interactive, and it stays on your plan. If Claude is running without anyone watching, on a schedule, in a GitHub Action, or behind an editor that talks to the API on your behalf, that is agentic, and it draws from the credit. Most teams find one or two background jobs they had genuinely forgotten about: a nightly documentation pass, a test-writing agent wired into continuous integration, a bot that summarises tickets. Each of those is fine. The point is to know they exist before they start billing separately.
What Australian teams should do before 15 June
The risk here is not the credit itself. It is the automated traffic you forgot was running. A few checks now will save a surprise invoice later.
Audit where your usage flows: interactive CLI and chat stay covered, but editor traffic routed through ACP and any CI jobs do not. Map which of your workloads sit on which side of the line.
Keep editor work on the subscription where you can: if you run Claude Code inside Cursor or Zed, running the official claude CLI in the editor's terminal still draws from plan limits rather than the credit.
Price your agentic workloads: a team running nightly continuous-integration agents could exhaust a Pro credit worth about $30 AUD in days, while the roughly $300 AUD Max 20x credit suits sustained automation.
Choose your overflow behaviour deliberately: enabling API-rate overflow gives continuity at a variable cost, while a hard stop caps spend but can pause your agents mid-month.
Set usage alerts now: configure them before the change lands, not after the first unexpected charge.
For an Australian SMB paying for a handful of Max seats, the realistic worst case is not the monthly credit. It is unplanned API overflow from a single agent loop left running. An unattended job can cost more across a weekend than the subscription costs for the month. The teams that come through 15 June without a shock are the ones that know exactly which of their Claude jobs run on their own, and have a cap or an alert sitting on each one.
Sorting interactive from agentic usage, pricing each workload and putting the right controls in place is a half-day job done once. If you would rather have it mapped for you, Automata AI helps Australian teams structure Claude plans, agent workloads and cost controls so the 15 June change is a line item, not a shock. You can book a short call here.
