Claude's Billing Changes: What Breaks, and How to Keep Your AI Agents & Automations Free
Anthropic is splitting Claude usage limits & it impacts our AI Agents. What changes, what stays free (Cowork is safe), the Enter Key Test to know if you are affected, and how to keep automations free.
If you opened Reddit or X this week, you'd think Claude just got twice as expensive overnight. The threads are loud. One developer posted that an agent ran up "around 50 bucks" building a single feature over a long weekend.
Another thread, with more than a thousand upvotes, is titled "Agent Use is gonna drop off a cliff once it's all usage based." Somebody already shipped an open-source tool to dodge the change.
The panic is real, and it's mostly aimed at the wrong people. Here is the grounded version.
I run automations on Claude every day, including the Content Radar AI Agent that I will ship next week for you. So I read Anthropic’s actual support document instead of the hot takes, checked it line by line, and wrote down what is true, what is rumor, and what to do. Let me walk you through it.
On June 15, 2026, Anthropic is splitting how Claude is billed into two pools. For almost everyone reading this, nothing changes unless you build AI Agents and Apps that harness Claude Code subscription keys.
For the AI builders group running any specific kind of automations and AI Agents, a few things change, and you have 18 days to handle it without paying a cent more.
👋 Julley, I'm Dheeraj, an AI systems builder.
I build production-grade AI systems at work by day and ship my own products by night, 9 and counting, including SubflowAI and the Content OS Agents Toolkit. This newsletter is the bridge between those two worlds. Every system, every build, documented step by step.
Join 1,800+ builders getting the exact AI setups, prompts, and workflows that actually work in your business.
The Claude billing change doesn’t touch the way most of us use Claude every day. It touches the AI Agents we set up to use Claude for us. Knowing which group you’re in takes about ten seconds.
What is actually changing with Claude usage limits on June 15?
Starting June 15, 2026, Claude Agent SDK usage and the claude -p command (the headless, non-interactive way of running Claude Code) stop counting against your normal Claude subscription limits.
Instead they draw from a separate monthly Agent SDK credit:
$20 on Pro,
$100 on Max 5x,
$200 on Max 20x.
Your subscription limits themselves do not change. This is a carve-out for automated usage, not a price increase on the plan you already pay for.
Think of it as Claude drawing a line down the middle of your account.
On one side sits everything you do by hand:
chatting on the web,
typing into Claude Code,
working inside Cowork.
That side keeps running on your subscription exactly as before.
On the other side sits everything a script or a scheduled job does for you while you are not watching. That side now pulls from the new monthly credit.
Four things move to the new credit pool, per Anthropic's support doc:
Claude Agent SDK usage in your own projects (Python or TypeScript)
The
claude -pcommand in Claude Code (the non-interactive, headless mode)The Claude Code GitHub Actions integration
Third-party apps that log into your Claude subscription through the Agent SDK
If none of those phrases describe anything you have set up, you can stop worrying right now. The next section explains why.
Does the June 15 change affect Claude Cowork?
No. Claude Cowork is explicitly excluded from the new credit and stays on your normal subscription limits, including its scheduled tasks.
Anthropic's support document names Cowork directly: "Your subscription usage limits stay the same and stay reserved for interactive use of Claude Code, Claude Cowork, and Claude."
So your Cowork morning tasks automations, your folder-based AI agents, and your connectors keep running exactly as they do today.
This is the part the developer threads skip, because most of them are not Cowork users. For our audience it is the headline. Cowork is the no-code automation layer most of you are building on, and Anthropic just confirmed in writing that it is untouched.
The Content Radar AI agent that I run every day and I will publish next week for you, for example, runs perfectly fine inside Cowork after June 15 with no change at all.
The same goes for interactive Claude Code (you, in a terminal, typing prompts) and every Claude chat on web, desktop, and mobile. None of it moves. None of it costs more.
Commercial break: Claude Code Builder cohort
The founding batch of my Claude Code cohort starts in July on Maven. Four weeks, six sessions. Learn to build the AI system with me.
Only 12 Seats. When they’re gone, the founding price ($797) closes and Cohort 2 opens at $1,597.
Use code GENAI20 for 20% off. Expires July 9. Check the Syllabus →
Anthropic increases Claude usage limits
It helps to know why this is happening, because the headline makes it sound like a takeaway when it is closer to the opposite.
Anthropic actually raised interactive usage limits this month even though personally I do not see any change other than lesser output than previously. If you work hands-on or interactive mode, you have more runway than you did in April, not less as per their claim.
The split targets the other side of the line: people who were running thousands of dollars of automated API value through a single flat subscription, and this carve-out separates that programmatic usage out so it gets metered on its own.
Hands-on use got more generous (again in claims). Hands-off automation got its own meter. That is the whole story, minus the panic.
Because of all this frustration I have created a backup agent agent harness with Codex that easily hands over the tasks from Claude to it or vice-versa when my workflows are hit with usage limits on either side. I will cover that in another article soon.
How do I know if my automation is affected?
Use what I call the Enter Key Test. If a human has to press enter for the work to happen, you stay on your subscription. If a robot presses enter for you while you are away, that usage moves to the new metered credit on June 15. That single distinction sorts almost every case correctly.
To be sure, answer three yes-or-no questions:
Do I have a script, cron job, or scheduled task that runs Claude without me sitting there typing? (That is
claude -por headless mode.)Did I build or buy a tool that logs into my Claude subscription and runs Claude in the background? (That is the Agent SDK or a third-party app.)
Do I have a GitHub Action that runs Claude Code automatically on commits or pull requests?
If you answered no to all three, you are not affected. Keep using Claude exactly as you do now. If you answered yes to any of them, you are in the small affected group, and the rest of this guide is your 18-day plan.
What happens when the Claude Agent SDK credit runs out?
The credit is measured in dollars, refreshes monthly, does not roll over, and drains before any other source. When it is gone, one of two things happens, and you control which.
If you have enabled pay-as-you-go usage credits, additional automated usage continues at standard API rates.
If you have not enabled them, your automation simply stops until the credit refreshes next cycle.
By default there is no silent overage. Your bill does not quietly balloon unless you turned that option on yourself.
That last detail matters for anyone who has ever been "scared of racking up a bill," which is one of the most common worries I hear in my DMs. The default behavior is a hard stop, not a surprise invoice.
You have to opt in to spend beyond the free monthly credit, which I do not recommend without proper guardrails.
Claude API credits don’t switch automatically
One thing not to miss: the credit is not switched on for you automatically.
You have to opt in once to activate it, and Anthropic is expected to send the activation email in early June.
If you have an affected automation, watch for that email and do not let it sit unread, because until you claim the credit your claude -p and Agent SDK calls have nothing to draw from.
The credit is also per person, not shared across a team, and it does not roll over month to month.
So for most affected solopreneurs, the real question is simple. Will my automation's monthly usage fit inside the $20 Pro credit or $100 Max plan credit, or will it run dry partway through the month?
Answering that is move number two in my Claude usage limits playbook below.
How much will this actually cost if you are affected?
It depends entirely on how hard your automation works, and the gap between the two ends is enormous. Be honest about which end you are on, because the credit is generous for one and small for the other.
If you run a light automation, a daily summary, a weekly research pass, a once-a-day brief, you are almost certainly fine. A single modest run is a few cents to a couple of dollars.
My own Content Radar AI agent measures $0.65 per run. Run that kind of job once a day and you will use most of the $20 Pro credit.
If you run a heavy, always-on AI agent, the story flips.
A serious agent that wakes up, plans, then spawns several workers that each call Claude many times to read files, use tools, and produce a result can fan a single job out into twenty or thirty model calls.
When some of those calls reach for the most expensive model like Opus 4.7, one overnight (couple of hours) run can cost several dollars. At one run a day, you can drain even the $200 Max 20x credit in a couple of weeks only, and then every extra call bills at full API rates or stops.
This is the part the "$200 dollars is free money" takes leave out.
The dividing line is simple. Light, scheduled, single-purpose jobs fit inside the credit comfortably. Always-on, multi-step, model-hungry agents do not. My playbook below works for both, but the heavier your agent, the more it matters.
Common mistakes people are making
A few avoidable errors or rumors are spreading along with the panic. Skip them.
Believing the Microsoft rumor. A widely shared claim says Microsoft canceled its internal Claude Code licenses over token billing. That traces back to a single social post, not to Microsoft or Anthropic. Treat it as unverified gossip, not a reason to act.
Assuming the whole plan got metered. Only automated SDK and headless usage moved. Your interactive use did not. People are rushing to “fix” things that were never affected.
Reaching for clarp first. The gray-area workaround is getting attention because it is novel, not because it is wise. Exhaust the three sanctioned options before you even consider it.
Turning on usage credits without meaning to. If you enable pay-as-you-go overage, an automation that drains its credit keeps spending at API rates. Leave it off unless you have a deliberate reason, so your default is a hard stop.
Get PluggedIn
The free half of this guide tells you whether June 15 even touches you. The paid half is the exact plan that makes sure it costs you nothing.
Skip it and the risk is the quiet kind. One automation you forgot about flips to full API rates on June 15 and quietly bills you for weeks before the charge ever catches your eye.
Your PluggedIn asset for this post
The June 15 Automation Audit Checklist, a fill-in worksheet you can run tonight:
The Enter Key Test up top, so you can sort any automation in ten seconds
A copy-paste audit table to list every place Claude runs and flag which ones change on June 15
A burn-estimate worksheet to see whether each automation fits inside your 20, 100, or 200 dollar credit
Four practical workarounds to keep your Claude automations and agents running FREE, all ranked from safest to riskiest (including the one that can get your account flagged, with an honest warning), and a build-in-public look at which lane I am choosing for my own agent and why.
A before-June-15 checklist including the easy-to-miss opt-in email
Hand the whole sheet to your Claude and say “help me fill this in,” or work through it by hand
The 3 moves to make before Claude billing changes on June 15
Everything in the affected group comes down to three moves. Do them in order. The whole thing takes an afternoon at most, and for many of you it takes fifteen minutes.
Move 1: Audit where Claude runs without you
List every place Claude runs that a human is not actively driving. Open your scheduled tasks, your cron jobs, your launch agents, and any tool you connected to your Claude account.
For each one, ask the Enter Key Test question: does a person press enter, or does a machine?
Here is the copy-paste checklist I used on my own stack. Run through it line by line.
