On June 15, 2026, Anthropic sent its subscribers an email. As of that day, usage of the Agent SDK, claude -p, and the third-party apps riding on top of them was supposed to be carved out of the subscription allowance and moved onto a separate monthly credit. The change had been announced a month earlier. And yet the same email said it wouldn’t be going ahead — not long after the announcement, not on some other day, but on the very day it was due to take effect. “Nothing changes for now,” the message read.
Overshadowed by the same company pulling its top model, Fable 5, entirely offline three days earlier under a US government order, this small retreat drew less notice. If switching off the model was a matter of nationality and security, this episode touches something quieter but more structural: how on earth you’re supposed to price the use of AI, and why today’s flat-rate model has such trouble doing it.
A new kind of customer called “programmatic use”
First, look at what was being carved out. The key word is programmatic use. If interactive use is a person sitting at the keyboard, asking one line at a time and reading the reply, programmatic use is code and automation calling the model in the person’s place. Automation apps (tools built with the Agent SDK), headless commands running in a terminal (claude -p), jobs firing automatically in a CI pipeline, and external agent tools all fall here.
The difference is speed and rest. A person stops to think, sleeps, and is bound by the natural ceiling of finger speed — a few dozen requests a day at most. An automated agent, by contrast, doesn’t stop. It calls the model hundreds or thousands of times within one task, retries on failure, and loops its own output back as its own input. Even under a subscription in one person’s name, the texture of consumption is completely different.
Anthropic’s original plan was to lift this programmatic use out of the subscription pool and move it onto a separate, plan-tiered credit. According to one report, that credit ran about $20 a month on the cheaper plan and around $200 on the higher one — drawn down at API list price, with nothing rolling over to the next month. Why attempt this split at all? The answer lies in a weakness of the flat-rate model itself.
Flat rates price the “average”
There’s a reason a monthly flat subscription feels cheap: the provider prices it for the average user.
A gym membership makes it easy to see. A gym can sell a monthly pass at an easy price because it knows most people who sign up won’t show up every day. If every member occupied the equipment for two hours without missing a day, that price wouldn’t hold. A buffet is the same: the entry fee is built on the average that “a typical guest eats about a plate and a half.” A flat rate is, in essence, a model where the light users’ slack covers the heavy users’ overflow.
Software subscriptions thrived on this model for twenty years on one decisive condition: marginal cost near zero. Marginal cost is what the provider pays extra when one more person uses the thing one more time. Installing already-built software for one more person, or playing an already-uploaded video one more time, costs essentially nothing extra. That’s what made the word “unlimited” possible — unlimited storage, unlimited streaming, unlimited use. Even when one person used far more than the average, the flat rate held because that overflow was close to free.
There’s a usage cap — so why the loss?
A natural objection: “Don’t subscriptions have a usage cap? If I only use what’s inside the cap, where’s the loss?” Fair point — and exactly where the heart of it is.
A subscription’s usage cap — the ceiling on calls per minute or per hour — is a ceiling, not the basis of the price. The price, as noted, is set to average use; the cap exists separately, as an upper bound to prevent abuse. And that ceiling is usually set generously enough that even an interactive heavy user — say, a developer writing code all day — never feels cramped. Set the ceiling right at cost, and ordinary users get squeezed.
The problem is that a person can’t fill this generous ceiling, but automation can.
A person’s consumption plays out well below the ceiling. Automation’s consumption bangs its head against the ceiling and tops itself back up every time the cap resets. The result is that under the same ceiling, the monthly running total diverges sharply. Among developers, the math made the rounds: convert a heavy user’s ceiling-scraping usage to API list price and it easily runs to several times the subscription fee. The ceiling itself was already designed to permit far more value than the fee.
This is where the premise of the flat rate breaks. LLM inference isn’t like copying software. Every additional token the model spits out costs real compute — GPU time. Marginal cost isn’t zero. So the more heavy users there are, the more the provider’s cost swells right alongside. The structure of patching overflow with light users’ slack doesn’t hold up against an agent that never rests. The cost one user runs up exceeds what they paid, and the more they use, the deeper the loss runs. The per-user balance — the unit economics — flips.
And there’s one more hole. When a third-party tool draws on the model with a subscription’s auth token, it becomes possible for several users or services to share — or effectively resell — a subscription under one person’s name. One person’s fee drains compute for several people’s worth. That’s the same reason Anthropic first poked at this hole back in January, when it blocked subscription tokens from being used by external tools.
The math checks out — but they couldn’t ship it
By this point, the decision to carve out programmatic use and move it to usage-based billing looks economically natural: charge for what’s used, at cost. And yet Anthropic stopped on the very day of rollout. Because price isn’t set by economics alone.
For one, there was competition. According to reports, around this same time the rival OpenAI was actually weighing a price cut. When one side is dropping prices and the other forces through what amounts to a hike, you hand price-sensitive developers a reason to switch. In the middle of a price war, flipping to usage-based billing is close to an own goal.
The business timing was bad, too. Anthropic was reported to be preparing an IPO, and for a company heading into a listing, there’s no worse signal than customer churn. In fact, some enterprise customers who saw their bills jumping from a flat rate to thousands of dollars a month were said to have already started cutting spend.
Above all, user trust was wobbling. The developer community pushed back hard, and a company post framing the change positively drew rebuttal annotations almost immediately. On top of that, just three days earlier, the US government had ordered the same company to block Fable 5 and Mythos 5 from foreign users, and the company had switched both models off for everyone that day. On one side the government forcibly shuts off a model; on the other the company changes how it charges. When the two land in the same week, what’s left for the user is a single feeling: I don’t control the AI I use. When you’re borrowing, neither the power switch nor the price tag is in your hand. Forcing through a price hike in a week when trust is shaking like that is not a good move.
What’s interesting is that this isn’t the first time the company has backed off the same problem. In January it blocked external tools from using subscription tokens, retreated under pushback, then in February amended its terms and wrote the same restriction into the contract after all. Make the aggressive move, retreat under pushback, then push it through again in a changed form. This June’s deferral reads naturally as part of that same arc. Even the email’s last line wasn’t a cancellation but a heads-up: when they change it again, they’ll let you know in advance.
Where “unlimited” ends
There’s a reason this doesn’t look like just one company’s pricing hiccup: the same pressure bears on generative AI as a whole.
The golden age of software pricing over the last twenty years stood on a foundation of near-zero marginal cost. On top of it, the user-friendly promises of unlimited and flat rates were possible. Generative AI clears that foundation away. Every sentence generated, every image drawn, every few seconds of video rendered actually burns compute. Marginal cost is distinctly greater than zero. So the further you go toward image generation, video generation, and above all the agent that never rests, the more “unlimited” recedes and usage-based — pay for what you use — comes back.
That doesn’t mean flat rates will vanish. Users want a predictable fixed fee; providers want a variable fee tied to cost. The two don’t mesh easily. For a while what we’ll see is repeated pricing experiments trying to draw a line somewhere between them — a quiet usage cap laid over a flat rate that looks unlimited, a base subscription with metered credits folded in, or automation alone carved out and metered separately. Anthropic’s attempted split was one of these; the timing was just bad.
On June 15, the same day, Anthropic took back a change it had announced a month before — not because the math was wrong, but because the rivals, the listing schedule, and the developers’ pushback were more urgent than the math. The company was clear that this was a stay, not a cancellation. The fact that automation eats away at flat rates stays right where it was, and the bill it deferred will show its face again in the next email.
— tomte
Read next · Writing for a Web That Stopped Clicking: Zero-Click, HTTP 402, and GEO · You Can Outsource Thinking, But Not Understanding