Anthropic is putting its top Claude model behind a meter
Starting July 12, all three paid Claude tiers will pay per token on top of subscription. It is the first frontier AI lab to meter a consumer chatbot, not just its API.
Starting July 12, all three paid Claude tiers will pay per token on top of subscription. It is the first frontier AI lab to meter a consumer chatbot, not just its API.
Anthropic will start charging its Claude subscribers a separate per-token fee for using Claude Fable 5 beginning July 12 at 11:59 PM PT, three days from now, on top of the flat $20, $100, or $200-a-month subscription they already pay. Claude Fable 5, the consumer release tied to the Mythos 5 model Anthropic announced this week, is, according to Wired, the first frontier AI model gated behind usage-based billing in a consumer chatbot product rather than only at the developer API.
Until July 12, subscription covered the cost of running Claude's smartest model. After, anyone who wants to send prompts to Claude Fable 5 will owe a per-token fee on top of subscription: $10 per million input tokens and $50 per million output tokens, the same rates Anthropic publishes for its developer API, according to Wired. The three paid plans keep their existing features, including the rest of the Claude lineup.
That is a different shape than the API-vs-subscription split consumers have seen before. Until now, frontier chatbot pricing has lived in two clean lanes: developers pay per token through the API; consumers pay a flat fee that includes everything inside the chatbot. Add-on costs for image generation or extended reasoning have existed for years, but the core model itself has stayed inside the subscription. Anthropic is collapsing the two lanes at the top: subscription buys access to Claude, and compute on the smartest model is sold by the token in the same checkout.
A flat monthly fee gives a consumer predictable budgeting and the freedom to chat without watching a cost counter. A per-token layer on top reintroduces exactly that counter, and it gives Anthropic a way to monetize the heaviest users without repricing the bulk of its subscriber base. Whether heavy users accept the meter or downgrade to the previous generation is the question that turns this from a pricing announcement into a business-model test.
Wired's worked example: a $20 subscriber who runs one million input tokens and one million output tokens through Claude Fable 5 in a month would owe an extra $60 on top of subscription, $80 total. Light users will see bills close to zero on top of subscription. Heavy users can climb past $200 a month before hitting the ceiling on the top tier.
The flat subscription model works when the average user costs less than the subscription fee to serve; labs price for that margin. A per-token layer on top implies that the average frontier-model user is now expensive enough to serve that heavy users should pay marginal cost, not just the average. That is the same reasoning that pushed the developer API to per-token pricing years ago. Consumer AI is catching up, in price if not in product.
Anthropic describes the release as a promotional access window for Mythos 5, set to expire when the meter flips on. The support documentation describes how Fable 5 access rolls into the new billing structure, a deliberately temporary wrapper around what amounts to a permanent design choice. Whether Anthropic extends, replaces, or removes the meter after the promotional period is one of the watch items the announcement leaves open. The clean test: if the metered tier survives Mythos 6, it is the new template. If not, it was a promo.
OpenAI's GPT-5.6 "ultra" tier, published the same day, raises the ceiling on the subscription side rather than carving out a separate metered lane. Anthropic is testing a different shape: keep the subscription, meter the smartest model on top, and see which dimension of the consumer trade, predictability or ceiling, flexes first. The two labs are answering the same question from opposite ends.
The bet is that consumers who want the best model will tolerate itemized AI compute bills the way developers already do. If they do, every other frontier lab will eventually have a chart like this in their planning deck. If they do not, Anthropic will have learned in public that consumer chatbot unit economics diverge from the API in more than price.
The meter goes live Sunday. The data starts then.