Amazon retools Anthropic AI billing, moving from compute-hours to token-based pricing, report says

Amazon has reportedly restructured a portion of its partnership with Anthropic, adjusting how it is billed for access to the companies’ artificial intelligence work. According to a Yahoo Finance report, the change moves billing for some services away from paying based on compute time, and toward a token-based model that prices based on the amount of data processed or generated by the AI system.

In token-based pricing, “tokens” are the units of text and other information an AI model consumes and produces. For companies deploying large language models at scale, shifting billing to a token metric can change the economics, because cost then tracks how much the service is actually used rather than how long the underlying computing hardware is running.

The same report says Amazon disputes a competing claim that the adjustment would increase its costs. The dispute, as characterized by the report, suggests that the company views the revised structure as either cost-neutral or preferable to the prior compute-hour approach, even if outsiders estimated otherwise.

The report’s framing indicates that the change is not just a technical accounting tweak. In practice, how AI partners price access can affect downstream pricing and margins across Amazon Web Services, or AWS, because enterprise customers and internal teams often measure usage and budget AI systems in terms of model interactions, prompts, and generated outputs, which align more naturally with token usage than with raw compute time.

Amazon’s broader push in AI services has put pressure on vendors and infrastructure providers to make billing models transparent and aligned with customer consumption. While the report does not describe the specific scope of which Anthropic workloads are affected, the existence of a token-based component points to an effort to connect partner costs more directly to how AI products are metered.

Industry context matters because pricing mechanics can influence how quickly AI deployments scale. If token-based billing better reflects real workload patterns, it can reduce incentives to optimize around compute runtime alone and instead encourage optimization around prompt efficiency and output length. That can be relevant for large customers and for AWS services that rely on predictable unit economics.

As of now, details remain limited. The Yahoo Finance post referenced does not appear to spell out the contract terms, the exact Anthropic model(s) involved, the start date of the new billing method beyond the reported timing, or the degree to which Amazon expects the change to alter total AI spending. It also does not provide the specific basis for the cost-increase claim that Amazon challenged.

Going forward, investors and customers will likely watch whether AWS and its AI customers see changes in how usage is billed, how internal AI unit costs are managed, and whether Amazon provides more clarity in earnings materials. Any later disclosure that quantifies impact, even in general terms, would be important given how sensitive AI margins are to inference (the process of generating responses) costs.