Anthropic launched its most capable consumer AI model, Claude Fable 5, on June 9, 2026. Three days later, the US government pulled it offline citing national security concerns tied to a reported method of bypassing its safety filters. The ban lasted 19 days. Fable 5 came back on July 1. Now, on July 12, it becomes the first flagship consumer AI model to sit behind pay-per-use billing on top of a monthly subscription.
These three events, in rapid succession, tell you a lot about where Anthropic stands right now.
The pricing shift is real and deliberate. Subscribers on any of Anthropic's plans, from the $20 basic tier to the $200 premium tier, will now need to buy usage credits to access Fable 5. The rate is $10 per million words sent to the model and $50 per million words it sends back. For a normal person having normal conversations, this will cost almost nothing extra. For someone running automated AI workflows in the background, the bill can climb fast.
This is not just about money. Anthropic is heading toward an IPO, with a Nasdaq listing currently targeted for October 2026. The company filed confidential paperwork with US regulators on June 1 and is valued at close to $965 billion. Yet it is still not profitable. It plans to spend around $19 billion on computing infrastructure this year. Bringing its highest-cost model under metered billing is a direct way to connect revenue to actual cost before public market investors start scrutinizing the books.
There is also a physical constraint driving this. Each new generation of Claude requires far more computing power to run than the last. Fable 5 is significantly more expensive per conversation than its predecessors, partly because it does more thinking before answering. Anthropic has signed deals with Amazon, Google, SpaceX, and Microsoft to expand capacity, but the bulk of that new infrastructure will not be ready until 2027. Charging by use is also a way to manage a scarce resource without simply turning people away.
The government episode added another layer of complexity. The US Commerce Department ordered Fable 5 offline on June 12, just three days after launch, over a jailbreak technique reported by Amazon researchers. Amazon is also one of Anthropic's largest investors, hosting the company on its cloud infrastructure, which made the sequence of events unusual. The ban was lifted on June 30 after Anthropic trained a new safety filter and made commitments to regulators. The episode proved something important: a government can switch off a commercially deployed AI model for all users globally with a single administrative order.
For businesses relying on Fable 5, that is a meaningful operational risk to register. It does not mean avoiding the tool. It means having a backup plan, because the last 19 days showed that even a fully paid, fully operational subscription is not a guarantee of continuous access.
The broader trend is clear. The AI industry spent 2024 and 2025 acquiring users with flat-rate pricing. That phase is ending. Anthropic started charging its largest business customers based on actual usage last year. AI coding tools like Cursor made the same shift. The logic is simple: AI agents, which can run for hours completing tasks automatically, consume far more computing power than a person typing questions. Flat fees designed for human-paced conversations do not survive contact with machine-paced automation.
For the average professional using Claude to draft documents, summarize reports, or answer questions, the usage-based fee for Fable 5 will likely add very little to the monthly bill. For teams building automated processes on top of Claude, the cost structure just changed materially, and budgeting for AI should now include a variable line item rather than just a fixed subscription.