Amazon engineers are reportedly trying to lower the cost of using Anthropic models by distilling them into smaller versions for internal use. The reported work points to a practical tension inside large AI partnerships: access to powerful models can be valuable, but the way that access is priced can quickly become a central engineering concern.
What Amazon engineers are reportedly doing
According to a report from The Information, some Amazon engineers are already distilling Anthropic models. The goal is to build smaller, cheaper versions that can be used internally.
Distillation is a model-development approach in which a smaller model learns from the outputs of a larger model. In this case, the larger models are Anthropic models, and the intended result is a version that is less costly to run while still benefiting from what the larger model produces.
The Information reported that Amazon has certain rights to use Anthropic's models for this purpose, according to a person familiar with the matter. The arrangement is described as similar to Apple's arrangement with Google Gemini.
The work matters because it suggests that model performance is not the only factor shaping enterprise AI decisions. Cost structure, deployment needs, and internal usage patterns can influence whether a company relies directly on a partner's model or tries to create smaller versions for specific use cases.
Why token-based pricing changes the calculation
The reported effort is tied to a renegotiation of the partnership between Amazon and Anthropic, according to The Information. Starting next year, Amazon will pay for Anthropic's models based on tokens processed rather than compute hours.
That shift could push costs up sharply, according to the report. A token-based model links the price more directly to how much the models process, while compute-hour pricing is based on the computing time used.
An Amazon spokesperson pushed back on the concern, saying the changes from the expanded partnership will not raise costs. Anthropic, for its part, points to lower prices relative to the performance its models deliver.
Those responses show that the same pricing change can be viewed in different ways. Engineers may focus on internal exposure to rising usage costs. The companies involved may emphasize the broader value of the partnership and the price-performance balance of the models.
Where Bedrock fits in
Amazon already offers a distillation service on its Bedrock cloud platform. But that service does not currently include Anthropic's Claude models.
Instead, the Bedrock distillation service supports Amazon's own Nova models and Meta's Llama models. That distinction is important because it separates Amazon's public cloud tooling from the reported internal effort involving Anthropic models.
In other words, Amazon has a distillation service, but the specific Anthropic-related work described by The Information is not simply a case of Claude models being available through that Bedrock feature. The report describes engineers using certain rights Amazon has to work with Anthropic models for this purpose.
That creates a clearer picture of the situation: Amazon is not only offering AI infrastructure to customers, but also managing its own internal model economics as its partnerships evolve.
Alternatives are also on the table
Amazon is reportedly exploring alternatives including OpenAI and its own Nova models. That does not mean Amazon is walking away from Anthropic. The source article also states that Amazon has invested up to $25 billion more in Anthropic this year.
At the same time, Amazon has invested up to $50 billion in OpenAI this year. Those figures show that Amazon's AI strategy is not limited to a single model provider or a single internal model family.
For Amazon, using multiple options could give teams more flexibility when they choose between Anthropic models, OpenAI, and Nova models. The reported distillation work fits into that wider pattern because it gives engineers another route: adapting the output of larger models into smaller, less expensive systems for internal needs.
The broader takeaway
The report highlights a simple business reality behind advanced AI deployment. Even when a company has access to high-performing models, the total cost of using them at scale can shape technical choices.
Model distillation is one response to that pressure. It can let teams preserve some of the value of larger models while seeking cheaper versions for routine or internal tasks. The source does not say how far Amazon's reported work has progressed, but it does show why engineers would be paying close attention before token-based pricing begins next year.
The debate is not just about whether Anthropic models are useful. It is about how Amazon expects to use them, how those uses will be billed, and whether smaller derived models can reduce the cost of internal AI systems.