Anthropic CEO Dario Amodei has suggested that AI companies could eventually charge different prices depending on the value or complexity of a task. His comments came in an interview on the Dwarkesh podcast earlier this year and have recently drawn attention online.

Amodei argued that not every token generated by a model is equally valuable. As examples, he contrasted routine advice such as telling someone to restart a Mac with guidance that helps a pharmaceutical company reorganize a molecule it is developing. He said the former might be worth a dollar or less, while the latter could be worth tens of millions of dollars.

He expects “pay-for-results” models to emerge, though he said it is too early to know which business model will prevail. Amodei also defended the current API model, saying it may last longer than some expect because developers can build on capabilities offered by future AI systems.

Anthropic’s S-1 filing says usage-based pricing accounted for about 80% of the company’s revenue in the first half of 2026, with subscriptions making up the remaining 20%. The filing also reports $16.3 billion in revenue for that period. A shift toward task-based or results-based pricing could change how AI services generate revenue, although Amodei did not describe a specific plan to introduce such a system.

The filing outlines substantial infrastructure commitments. Anthropic has more than $600 billion committed to future infrastructure, with the possibility of adding another $150 billion. Its total computing commitments are about $895.4 billion, including $70 billion due within the next 12 months.

The company’s revenue grew roughly 13 times to $33.2 billion over nine months, while it recorded a loss of $104 billion. The source report attributes most of that loss to a $94 billion accounting charge linked to financing agreements, citing the Associated Press. Amodei said in the interview that he considers the underlying economics profitable and that losses mainly reflect misjudgments about demand when buying computing capacity in advance.

Sources: Xataka.