You Can Still Run Agents on a Subscription (For Now)

We run an AI assistant that never sleeps. It sits on our own server around the clock, reading, replying, working on while we sleep. The day we set it up, the first question to answer was not about code. It was about money: how should it connect to Claude, through a flat monthly subscription, or pay-as-you-go through the API?

We almost picked the subscription. It looked simple and easy to budget, the same amount every month, no surprises. But one more layer of thinking gave us pause. The thing we were about to plug in was not us at the keyboard. It was a machine that works on its own, all the time. And those two use resources nothing alike.

A few weeks later, on June 15, 2026, Anthropic made an announcement that drew the same line for us, and it confirmed exactly what we had been uneasy about. How you pay for AI does not hinge on whether you use a little or a lot. It hinges on who is doing the using, a person or a machine.

Part 1It started with an agent that never sleeps

Most people use Claude by sitting and talking to it: open a chat or a terminal, type a question, read, type again. In a day you might ask a few dozen times. While you think, while you type, while you go eat, the machine sits idle. This kind of use has natural pauses built in.

An agent is not like that. It calls the model on its own, in rounds: read, think, call again. No break to go eat. The assistant we set up can run all night with nobody watching. A single task might call the model dozens of times back to back. This is why costing out an agent is a different exercise from costing out a person sitting and using it themselves.

Part 2Flat-rate vs pay-as-you-go: what each is for

First, be clear that these two bill on different logic, not just a different price.

Claude's subscription comes in tiers, from Pro at around $20 a month up to Max at $100 and $200. Its value comes from one thing: if you use it steadily, the flat rate beats paying per token. And what keeps the plan from losing money is the rate limit, the ceiling that stops you periodically. Put another way, the subscription can be a bargain precisely because it has a ceiling.

Now take an agent that never stops and put it on a plan that is a bargain because of its ceiling, and the contradiction shows. A machine running all day hits the ceiling fast and stalls, or else you end up using far more than the flat rate was meant to cover, which is not a business model anyone leaves running for long.

Part 3Right now, agents run on the subscription

Here is what many people miss: right now you can run agents on the subscription, not only when you are typing yourself.

Back in June 2026, Anthropic announced it would split the two kinds of use apart. From the 15th, anything calling through a script or an agent, the Agent SDK, the claude -p command, or outside tools wired in, would stop drawing from the subscription and move onto a separate credit priced at full API rates (Pro $20, Max $100, Max $200, no rollover to next month). Human use like chat or Claude Code in the terminal would be untouched.

But then Anthropic paused that announcement. In their own words, for now nothing has changed: the Agent SDK, claude -p, and third-party tools still draw from the subscription as before. The separate credit is not live, and they said they will share an update before anything takes effect.

In plain terms, the door is open today. Both the side where you type yourself and the side where an agent calls through the SDK run on the subscription. That is a window that works right now.

Part 4But the door may close: how we set it up

This window is not free forever. Announcing the split and then pausing it still tells you the direction: Anthropic wants to stop subsidizing heavy machine use on the flat plan. A single agent can fire the model thousands of times, rerun tests, call itself in a loop, a different scale from a person asking a few dozen times a day. One day the separate credit will likely come back in some form.

If a person is the one using it, look at the subscription first. If a machine stands in for a person around the clock, plan for the API.

So in practice we split agents into two kinds and put them in different places.

The first kind is an agent you drive through Claude Code on your own machine. The broad method: we do not call the API separately. We keep Claude Code open, logged in with the subscription account, and use that session as the parent, letting it fan out into sub-agents running in parallel under the one session. Everything runs under the session you are sitting at, so it counts as human use and stays on the subscription with no separate API bill. The dev fleet we build with every day runs exactly this way, on the plan we already pay for.

The second kind is a body that stands on its own 24/7, with no human session held open, like our assistant that never sleeps. There is no session for it to ride, so we connect it straight to the API from the start. Not because it cannot run on the subscription today, but because we do not want something that has to stay up forever hanging on a window the platform has already said it wants to close.

One more line worth stating plainly. Driving Claude Code interactively is what Anthropic itself says is unaffected, because it is human use. But feeding your subscription credentials into a headless runner, through an outside routing tool, to dodge API costs runs against the direction they announced. So we took the path that matches the plan's intent rather than the one that evades it.

Takeawayand where to start

One sentence to keep: match how you pay to who is doing the using, a person or a machine. Do not put an agent that runs all day on a plan priced for human hands.

1. Ask first who drives the work, a person or a machine. If a machine, does it run on a session you keep open, or stand on its own?
2. Drive it through Claude Code logged in with your subscription, letting it fan out sub-agents under one session, and it still runs on the plan today. Use the window, but know it may close.
3. A body that stands on its own around the clock, headless, with no session to ride, connect the API and pay per use, and set a budget ceiling so the bill does not balloon.
4. Do not feed subscription credentials into a headless runner through a routing tool to dodge API costs. That runs against the direction Anthropic announced.

Prices will keep changing. The who-is-at-the-keyboard rule will not. Start by asking that question of the work in front of you today.