Flat-rate AI is ending for agents. How to budget when every loop is metered
Anthropic's plan to move Agent SDK usage onto a separate metered credit was paused a month later, but the economics behind it are not going anywhere. A subscription that subsidised agents at 15 to 30 times the API price was never going to survive loops that run all night. Here is how to budget per task instead.
Update, 16 June 2026: Anthropic paused this change on 15 June, the day it was due to take effect, and says a revised plan will come with advance notice. The arithmetic below has not changed and neither has the direction. Read the credit amounts as the proposal that was on the table, and the budgeting method as the part that still applies.
On 14 May Anthropic told Claude subscribers that from 15 June, anything they ran through the Claude Agent SDK, the headless claude -p command, Claude Code GitHub Actions or a third-party app built on the SDK would stop drawing from their subscription limits. Those workloads would move to a separate monthly credit instead: $20 on Pro, $100 on Max 5x, $200 on Max 20x, $20 a seat on Team, billed at standard API rates, with no rollover and no overflow unless you switched extra usage on.
Interactive use stays where it was. Chat, Cowork and Claude Code in a terminal keep the same limits as before. The split is between a person typing and a loop running.
I think this was the most important pricing announcement of the year so far, and the fact that Anthropic pulled it back a month later does not change that. Somebody will do it, whether it is Anthropic in a revised form or a competitor first. The economics that forced the announcement are not going away.
The subsidy nobody put on a spreadsheet
Zed, whose editor connects to Claude through the Agent Client Protocol, published a note the day after the announcement. Their estimate, drawing on Matthew Diakonov's public cost comparison, was that a Claude subscription had been subsidising agent usage at roughly 15 to 30 times the equivalent API price. A Pro subscriber paying $20 could, in principle, drive several hundred dollars of API-equivalent compute through a loop and stay inside the plan.
That worked while most subscribers were people typing. A person sends a request, reads the answer, thinks for a bit, types again. A loop sends the next request the moment the last one returns, and it never gets tired. An agent that retries, re-plans and re-reads its context can burn in an hour what an attentive human uses in a week.
Every subscription business hits this wall once its heaviest users stop being humans. Anthropic had already banned OpenClaw-style harnesses in February and tightened the rules again in April, so the May announcement was the formal version of something the company had been enforcing by hand for months.
If your AI business case has a line that says "$200 a month, unlimited", it is built on that subsidy. Now is the time to find out what the number is without it.
What the announcement changes for a buyer
The change was narrowly targeted, and the targeting is the lesson. The workloads being moved to metered billing were exactly the ones a business would want to scale: scheduled jobs, CI pipelines, background agents, anything that runs without a person watching. The workloads left on the flat rate were the ones that scale with headcount.
So the question for a buyer is which of your AI usage is a person and which is a loop. Most teams I talk to cannot answer that on the spot. They have a handful of subscriptions, a couple of API keys, and a growing pile of automations that someone built on a Friday afternoon using whichever credential was nearest.
Sort it out. Anything unattended should be on an API key with its own budget, regardless of whether the paused change ever returns. You want to know what each automation costs on its own, and you want to be able to switch it off on its own.
Budget per task, never per seat
Seat pricing tells you what you will pay. It says nothing about what you are getting. For agent workloads the only budget line that means anything is cost per completed task, and you can build it from three measured numbers.
Tokens per task. Run a representative sample of the job through the agent and log input and output tokens separately, because output costs several times more than input on every major model. Include the retries. A task that succeeds on the third attempt costs three attempts.
Tasks per month. Take the real volume, then add the growth you are hoping for, because a successful agent gets given more work.
Failure rate. An agent that fails 15 per cent of the time and gets re-run costs 15 per cent more than its success rate suggests, and the failed attempts usually cost more each than the successful ones because they wander.
As an illustration only: an invoice triage agent that averages 40,000 input tokens and 3,000 output tokens per document, on a mid-tier model priced at $3 per million input and $15 per million output, costs about 16.5 cents per invoice before retries. At 5,000 invoices a month that is roughly $825, or closer to $950 once you allow for a 15 per cent retry rate. A finance team can put that against the cost of the two people who currently do the work. "Unlimited for $200" was never a number they could use.
Six things that move the number
Prompt caching is the biggest single lever and the least used. Anthropic prices cache reads at a tenth of the input rate. An agent that re-sends the same 30,000-token policy document on every turn is paying full price for it thirty times when it could pay once. If your agent has a static preamble, cache it.
Use the smallest model that passes your evaluation set for each step of the loop, and reserve the expensive model for the step that needs it. Planning might need the frontier model. Extracting a date from a PDF does not.
Cap the loop. Every agent should have a maximum number of turns and a maximum spend per task, and it should fail loudly when it hits either. An agent that cannot finish should not be allowed to keep trying until the credit card notices.
Log cost per task from day one, in the same place you log success and failure. If the two numbers live in different dashboards nobody will look at them together.
Give each workload its own credentials, so that one runaway job cannot eat the budget of the others.
And put the number in front of the person who owns the outcome, weekly, in a form they can read: cost per invoice processed, cost per ticket resolved, cost per lead researched. That is the unit they think in.
What to do this week
Audit the last 30 days of programmatic usage across every plan and key you own. Most teams find automations they had forgotten about.
Move anything unattended to its own API key and budget. Do this even though the June change was paused. A subscription is a good deal for a person and a poor way to run a fleet.
Switch caching on wherever the prompt has a static section. The gain is immediate and the code change is small.
Then rebuild your business case on cost per task. If it still holds without the subsidy, the project was real. If it does not, better to learn that now from your own spreadsheet than in eight months from an invoice.
We run this exercise as part of every AI strategy engagement, and we built the AI savings estimator so a team can get a first number without us.
Frequently asked
Did the Anthropic Agent SDK billing change take effect?
No. Anthropic announced on 14 May 2026 that Agent SDK, claude -p, GitHub Actions and third-party app usage would move to a separate monthly credit from 15 June. On 15 June the company told subscribers the change was paused and that a revised plan would be announced in advance. As of this update those workloads still draw from subscription limits.
Which Claude usage was going to be affected?
Programmatic use only: the Claude Agent SDK, the headless claude -p command, Claude Code GitHub Actions and any third-party app that authenticates through the Agent SDK, including editors connecting over the Agent Client Protocol. Chat, Claude Cowork and interactive Claude Code in a terminal were not part of the change.
How should a business budget for AI agent costs? Per completed task, from three measured numbers: tokens per task (input and output separately, retries included), tasks per month, and failure rate. Seat pricing hides all three. Add prompt caching, per-step model selection and hard caps on turns and spend, then put the resulting cost per task next to the cost of the manual process it replaces.
Related reading
- What production AI actually costs
- How to measure ROI on an AI investment
- AI agents vs automation: when to use which
The flat-rate era made agents look free, which was convenient for demos and useless for budgets. Metered agents are a better deal for any business that knows what a task costs, and a worse one for any business that does not. Find out which you are.