Anthropic’s eye-poppingly powerful new model, Fable, is worth testing while you still can. Built by the company behind the Claude chatbot, Fable is the publicly safe version of Mythos, the model Anthropic deemed too dangerous to release just two months ago. To make it available to the average user, Anthropic has introduced stringent guardrails designed to prevent the cutting-edge model from being used for high-risk applications.
Those who have used it describe Fable as a step change in AI capability, able to handle complicated tasks with ease. That’s one reason to adopt it—and quickly. But there’s another: In less than two weeks, it will disappear from most people’s budgets.
For now, Fable is accessible to monthly Anthropic subscribers on Pro, Max, Team, or Enterprise plans. But that’s only through June 22. After that, Anthropic is putting it solely behind API access, where users have to pay on a per-token basis. And if there’s one thing big businesses and everyday users are quickly realizing as their bank accounts are hit by ever-bigger AI bills, it’s that tokens aren’t free.
Anthropic says the limited subscriber access is due to capacity constraints. “As enough capacity comes online, we aim to make it a standard part of those [subscription] limits again,” wrote Anthropic’s head of growth, Amol Avasare, on X. “We’re sprinting as hard as we can at this,” he added. But Avasare said the company couldn’t make any promises on timing.
Increasingly, though, what looks like a temporary capacity hiccup feels more like an indication of AI’s future direction. Since the release of ChatGPT in November 2022, consumer AI has been built on a simple premise: Pay $20 or more a month, get access to the smartest machines on earth, and use them as much as the companies allow. Most users never hit their limits, while plenty of heavy users got effectively subsidized access.
But as agentic AI replaces chatbots as the newest way to interact with AI systems, the price of inference—the cost of delivering the service—has shot up. The better these systems become, the more likely people are to ask them to do longer, messier, more valuable work, even as the cost of the hardware needed to serve the models keeps rising.
That is why the industry is quietly moving from “all you can eat” to “eat what you can afford.” Last month, Google shifted Gemini toward compute-based limits; Microsoft’s GitHub recently cut back its usage limits; and at the end of May, OpenAI switched off its usage-limit multiplier after a monthslong trial.
All are examples of the endless AI subscription model being squeezed by the economic realities of delivering the technology. The “unlimited” frontier AI era was fun while it lasted.
