Anthropic Delays Extra Charges for Its Best Claude: 50% of Usage Kept, the Rest Extended Until August

Update (July 19, 2026): after several extensions of the announced timeline, the usage-based charges are not being applied as described in this article: subscribers keep 50% of their plan usage, and Anthropic has extended the additional 50% until August. We preserve the original analysis—with the announcement's figures and quotes—while flagging where the situation has changed.

Paying $20 a month for a chatbot and using it without watching the clock has, until now, been the norm. Anthropic announced it would break that norm — though, for now, it has postponed the move (see update). According to Wired, the company would require its subscribers—including those on its $20, $100, and $200 monthly plans—to pay additional usage fees to access its most powerful consumer model. It's apparently the first time a top-tier AI lab has put its best consumer model behind a meter.

The change goes beyond a simple price adjustment. It marks the moment the industry starts charging end users what artificial intelligence actually costs to run.

What the announcement said

Starting on a deadline announced by the company, subscribers who wanted to use its top-tier model would have to pay based on consumption, at the same rates Anthropic charges developers accessing the model through its API: $10 per million tokens sent to the model and $50 per million tokens the model generates in its responses.

A token is the smallest unit these systems use to chop up text: roughly three-quarters of a word. A million tokens works out to about 750,000 words—more than the entire Lord of the Rings trilogy. That sounds like a lot, and it is, for occasional use.

But the math balloons for heavy users. Wired offers a concrete example: a subscriber on the $20 plan who sends a million tokens in a month and receives another million back in responses would pay $60 extra—$80 total that month. To put that figure in context, $80 covers roughly five months of Amazon Prime.

And here's the detail that drives the bill up: newer models burn through far more tokens than it appears. Before responding, they run a hidden "chain-of-thought" process—a kind of internal scratch pad where the model reasons step by step. That scratch pad consumes tokens the user never sees but still pays for. It's not unusual for power users to rack up API bills in the thousands of dollars per month.

Where things stand today

The timeline, however, has not held. After several extensions, subscribers currently keep 50% of their plan usage, and the company has extended the additional 50% until August. The usage-based charges, as described, are postponed: the direction of travel remains, but neither the pace nor the initial scope survived contact with users.

Why Anthropic is doing this

The official reason is capacity. An Anthropic spokesperson, Reem Ateyeh, told Wired that the company intends to bring the model back to subscription plans "once sufficient capacity allows" and to do so "as soon as possible," referring to its compute constraints. Anthropic has signed multibillion-dollar data-center capacity deals in recent years with SpaceX, Amazon, and Google, and still says it isn't enough.

The underlying argument is that flat pricing simply no longer adds up. Nick Turley, former head of ChatGPT at OpenAI and now leading its enterprise products, put it bluntly on a podcast: "It's possible that, in this current era, having an unlimited AI plan is like having an unlimited electricity plan. It just doesn't make sense."

The logic gets sharper with AI agents—systems that carry out tasks autonomously, writing code or chaining together actions—because they consume far more computing power than a traditional chatbot. A user who only asks questions and one who sets an agent loose to work for hours pay the same under a flat rate, but they cost the company radically different amounts.

There's also a financial timing factor. Anthropic has already started billing large corporate clients based on how much AI their employees actually use, rather than charging a fixed fee. Wired notes these moves could be an effort to clean up its books ahead of a potential IPO.

Not an isolated case

The industry has been pushing in this direction for a while. Last year, AI coding startups like Cursor dismantled their unlimited plans and replaced them with usage-based pricing. Flat pricing worked while AI was a growth product subsidized to win over users; it stops working once profitability becomes the priority.

What sets Anthropic apart is that it's putting the meter on the end consumer and on its flagship model, not on some niche developer product. That's what turns this move into a real market test: is the average user willing to pay to always have the best?

The bigger bet: becoming the "Apple of AI"

Anthropic has focused mainly on the enterprise market, but it's increasingly encroaching on the consumer territory dominated by OpenAI's ChatGPT and Google's Gemini. The numbers explain why: according to analytics firm Sensor Tower, Claude hit 245 million unique visitors in a single month, more than double what it had just a few months earlier. That's still far behind ChatGPT's 1.11 billion or Gemini's 662 million, but it's growing fast.

The strategy, according to Wired, is to become the "Apple of the AI era": betting that there will always be a segment of users willing to pay more for the best. A source close to the company describes the shift in user mindset this way: the question is no longer "do I need the best intelligence for this task?" but rather "am I someone who needs mid-tier intelligence, or someone who needs the best?"

It's a risky bet. It depends on maintaining a technical edge, and the competition isn't letting up: the article itself notes that early reactions to OpenAI's new flagship model are complicating things for Anthropic.

The end of subsidized AI

The contrast with its rivals is the most telling part of this story. Neither OpenAI nor Google seems willing to follow the same path: both are expected to lean harder into advertising on their free and cheaper tiers to generate revenue. Anthropic, which has taken a hard public stance against advertising, has fewer options left. Raising prices may be, in the short term, its only real lever.

Both paths—ads or usage fees—point to the same conclusion: someone has to cover the real cost of running these models, and increasingly, that someone is the user. For years, cheap or free access to the best AI was propped up by investors willing to absorb losses in exchange for market share. That subsidy is running out.

For everyday users, the takeaway is straightforward. Anyone who uses AI occasionally will barely notice the change. But anyone who has built it into their daily work—developers, analysts, anyone who hands off long tasks to agents—will start facing variable, potentially steep bills, and will need to decide task by task whether the pricier model is really worth it. The comfort of a flat fee, with its all-you-can-eat feel, is giving way to something closer to a power bill: the more you use, the more you pay.

One question Anthropic itself leaves unanswered: if the constraint really is compute capacity, it's unclear when—or whether—that constraint will ease enough to bring the best model back into the subscription. In the meantime, the message to the industry is unmistakable. The era of giving away frontier AI for free to win users is over.