I was reviewing our team's CI/CD spend on a Thursday night when I noticed the number was off. Three GitHub Actions workflows I had built to run automated Claude code reviews were drawing more credits than I expected. Not a lot more. Just enough to make me open the Anthropic billing console and actually read the changelog.
That is when I found the notice about the anthropic billing split.
Starting June 15, 2026, Anthropic is dividing every Pro and Max subscription into two separate credit pools. One for interactive use. One for the Agent SDK. If you have automated pipelines calling Claude and you do not act before the changeover date, those pipelines will hit limits that did not exist last week.
This is what changed, who it affects, and exactly what to do about it before the deadline.
What the credit pool split actually means
Right now, one pool covers everything you do with Claude. You open the app and chat: draws from the pool. You run claude -p in a shell script: draws from the same pool. A GitHub Action invokes the Agent SDK at 3am: same pool, no distinction.
On June 15, that ends.
Interactive pool. Everything a human does in real time: the web app, the mobile app, the desktop client.
Agent SDK pool. Everything programmatic: the CLI, GitHub Actions, application code calling the SDK, scripts, background jobs, scheduled tasks.
The Agent SDK pool bills at full API list rates. There is no rollover. Interactive credits cannot cover Agent SDK usage. If your automated usage exceeds your monthly allocation, the overage is billed on top of your subscription fee at standard API pricing.
The allocations by plan tier:
- Pro subscribers: $20/month in Agent SDK credits
- Max 5x: $100/month
- Max 20x: $200/month
Standard API pricing is roughly $15 per million input tokens and $75 per million output tokens for Opus 4.8. For high-volume workloads, that math matters.
Who this affects
If you use Claude only through the chat interface, this change is invisible to you. Nothing changes.
You are affected if any of the following is true:
You run claude -p in shell scripts or terminal automation. You have GitHub Actions that call the Agent SDK. You have built internal tools that use the SDK to process documents, summarise tickets, or review code. You are using Claude Code in a CI/CD pipeline. You have scheduled jobs that invoke Claude on a timer.
The teams most likely to get caught off guard are those who started using Claude interactively, liked it, and gradually automated a few workflows without ever thinking of themselves as API users. Their consumption patterns changed months ago. Their mental model of "I just have a Claude subscription" has not caught up.
If someone on your team built a Claude integration as a side project and nobody tracked the usage since, June 15 is a good reason to go find it.
The napkin math
Let me work through a concrete example.
A small development team runs 10 automated code reviews per day. Each review handles roughly 2,000 input tokens and produces 500 output tokens. That is 20,000 input tokens and 5,000 output tokens per day. Scaled to 30 days: 600,000 input tokens, 150,000 output tokens per month.
At Opus 4.8 pricing: 600,000 input at $15/M = $9. 150,000 output at $75/M = $11.25. Monthly total: $20.25.
A Pro subscription just barely covers it. No overage.
Now scale that to 50 reviews per day, plus a nightly document processing job, plus a weekly codebase audit. You are probably looking at $100-150/month in Agent SDK usage. The Pro pool runs out in the first week. Max 5x ($100/month allocation) covers you. Max 20x ($200/month) gives you headroom.
Go above $200/month and the calculation flips: a dedicated API account with committed spend pricing starts to look cheaper than subscription tiers plus overage fees. The break-even point varies by usage pattern, but it is in the $200-300/month range for most teams.
The honest version of the napkin math: check your actual usage in the Anthropic console. The billing dashboard already shows the interactive versus programmatic breakdown. Look at the last 30 days. That number is your baseline.
Five things to do before June 15
One: log into console.anthropic.com. Navigate to the billing section and find "Agent SDK Credits." This is where you confirm your allocation and formally claim it. The credits may not be applied automatically on June 15 if you have not done this step.
Two: inventory your automated workflows. Make a list. GitHub Actions. Cron jobs. Internal tools. Background processors. Scripts that a colleague wrote six months ago and everyone forgot about. Write down everything that touches the Claude API programmatically. This is not optional. You cannot estimate your exposure without it.
Three: estimate your monthly token consumption. Pick two or three of your heaviest workflows. Look at their last 30 days in the usage dashboard. Scale up to your full list. You are answering one question: is your monthly programmatic usage closer to $20, $100, or $200?
Four: make the plan decision. If you are under $20/month, Pro covers you. If you are in the $50-100 range, upgrade to Max 5x before June 15. If you are above $100, upgrade to Max 20x or start migrating heavy-volume workloads to a dedicated API account. Do not wait to figure this out on June 16.
Five: set a spend alert. After you claim your credits and pick your tier, set up a spend alert in the console for 75% of your Agent SDK allocation. This is the part most people skip. An alert at 75% gives you time to investigate and adjust before you hit overage pricing.
What happens after June 15
If you have claimed your credits and your usage fits within your allocation, the transition is invisible. Your workflows run. The credits decrement. Nothing breaks.
If you have not claimed your credits, your Agent SDK pool starts at zero. Programmatic usage will be billed at full API rates from day one, charged directly to your payment method on file.
If your usage exceeds your allocation mid-month, you will not get a hard stop. The API keeps working. You just shift to overage billing, and the charges appear on your next invoice. The console will show you when you have crossed the threshold.
The error logs in your workflows will not look different. There is no specific error code that says "you hit your Agent SDK limit." The requests succeed. The bill just changes. This is the part that makes it easy to miss for teams that do not monitor their Anthropic spend closely.
Why Anthropic is doing this
This is not a punitive change. The split reflects a real operational difference between interactive use and agentic workloads.
Interactive users are bounded by human attention. They might run intensive sessions during work hours and be silent at night. The load is spiky and self-regulating.
Automated workloads are different. They run at machine speed. They can consume infrastructure continuously in ways a human session never would. Mixing them in the same pricing bucket was always going to cause problems as agentic usage scaled.
Other providers will do this. The distinction between "human using AI" and "software using AI" is real, and the billing structures across the industry will reflect it. Anthropic is moving first, and the announcement was quieter than it should have been for something that affects this many teams.
The frustration is not the policy. The frustration is the gap between how quietly it was announced and how broadly it affects people building on Claude.
For building efficient Claude workflows before the changeover, the Claude Code GitHub Actions CI/CD guide covers how to structure automated pipelines to minimise token consumption. If you are deciding between building on the Agent SDK versus a dedicated API account, the agentic engineering complete guide covers that tradeoff at the architecture level. And for teams building custom agents, the Claude custom agents guide includes patterns for estimating and controlling usage before you scale.
Frequently asked questions
Does this change affect dedicated API accounts? No. If you are already on a dedicated Claude API account with a separate API key and paying per token, nothing changes for you. This split applies exclusively to Pro, Max 5x, and Max 20x subscription holders who access the Agent SDK through their subscription.
What happens if I do not claim my credits before June 15? Your workflows will continue running, but your Agent SDK pool may start at zero if the credits are not formally claimed. The process takes about 60 seconds in the console. Do it now.
Can I transfer credits between the interactive and Agent SDK pools? No. After June 15, the pools are completely separate and non-transferable. Interactive credits cannot cover programmatic usage and vice versa.
What error will I see if I hit the Agent SDK limit? You will not see an error. Requests continue to succeed. The difference is that they shift to overage billing. Monitor your console spend dashboard rather than relying on error logs.
How do I see my current programmatic versus interactive usage split? Open the Anthropic console, navigate to Usage, and look at the detailed breakdown view. It already separates interactive and programmatic consumption. This is your baseline before June 15.
Should I move all automation to a dedicated API account now? If your monthly Agent SDK usage is consistently above $200, a dedicated API account with a committed spend tier will likely be cheaper than Max 20x plus overages. Below $200/month, subscription tiers are generally competitive. Run the math with your actual usage numbers.
Will other AI providers follow with similar billing splits? Almost certainly. The economics of interactive versus agentic usage are fundamentally different, and every major provider is building pricing models that reflect that distinction. Anthropic moving first does not mean it will be alone for long.
