The growing prominence of Anthropic as a mainstream provider of large AI models has strengthened its negotiating position with clients, including one of its earliest and most significant investors, Amazon.com.
According to two individuals familiar with the details of the negotiations, earlier this year Anthropic renegotiated parts of its agreement with the cloud computing and e-commerce giant. This will result in Amazon.com incurring higher costs for using Anthropic's models in its own products. One of the sources indicated that due to these rising expenses, Amazon is evaluating the use of alternative models to reduce costs, with options including models from OpenAI and Amazon's own internally developed Nova model.
Details of the New Pricing Model
Sources revealed that the new pricing scheme, set to take effect next year, will charge Amazon.com based on the number of tokens processed, which is a fundamental unit of information for AI. This replaces the previous model where fees were based on actual compute hours consumed.
This change in billing rules is likely to significantly increase Amazon's AI development expenses. Amazon has launched several AI products for both consumers and businesses that rely on Anthropic's technology, including the shopping assistant 'Alexa Shopping', the code development tool Kiro, and the workplace assistant Quick.
A spokesperson for Amazon.com stated, "Amazon and Anthropic have a deep, multi-faceted technical partnership, and we continue to maintain and deepen our collaboration. Claims that expanding our partnership will increase our costs are not accurate."
Although Anthropic recently adjusted its billing rules for Amazon and other customers, the company maintains that the overall cost of using its models is on a downward trend.
An Anthropic spokesperson said, "With each generation of the Claude model, the cost for users to accomplish core tasks continues to decrease. In November 2025, we significantly reduced the pricing for the Opus model, and that price has remained stable since, while the model's capabilities have continued to improve, delivering substantially more compute value for customers' budgets."
Strategic Investments and Rivalries
This pricing adjustment occurs as Amazon.com is strengthening its ties with Anthropic's primary competitor, OpenAI. Earlier this year, Amazon committed to investing up to $50 billion in OpenAI. As part of the deal, OpenAI will deploy its computing infrastructure on Amazon Web Services (AWS), AWS will distribute OpenAI's full suite of models to its customers, and Amazon gains the right to use OpenAI's AI technology in its own products. Additionally, Amazon finalized plans this year for an additional investment of up to $25 billion in Anthropic.
Recent actions by Amazon have further complicated the situation amid ongoing tensions between Anthropic and the U.S. government. CEO Andy Jassy reportedly warned White House officials about potential security risks in Anthropic's two latest models. This communication contributed to the administration implementing restrictions that barred non-U.S. persons from accessing the 'Mythos' and 'Fable' models, leading Anthropic to broadly suspend public access to them.
Anthropic officially stated that the partnership continues to deepen: "Amazon is one of our most important partners, with over 100,000 business customers already building AI applications using the Claude model on the AWS platform."
Token-Based vs. Compute-Hour Billing
The token is the standard unit of billing in the global large model industry. This year, Anthropic shifted its pricing model for most customers from a subscription plan with a monthly cap to a pure pay-as-you-go model based on usage, significantly increasing costs for many enterprise clients. Some companies have chosen to absorb the higher costs, while others have turned to open-source models to reduce expenses. Amazon was already on a pay-as-you-go plan, but its fees were previously tied to compute-hour consumption.
The commercial relationship between Amazon.com and Anthropic dates back to 2023, when Amazon announced a $4 billion investment in Anthropic. In exchange, Anthropic agreed to use AWS as its primary cloud provider and to adopt Amazon's custom AI accelerator chips, Trainium and Inferentia.
AWS not only provides the massive computing resources needed for training Anthropic's models but also grants access to a vast pool of enterprise customers. In 2023, AWS launched the Bedrock service, allowing business clients to access leading AI models from various providers in one place, with Anthropic being one of the first partners.
However, this collaboration has come at a high cost for Anthropic. Previous reports indicated that for every instance an Anthropic model is used via AWS, Anthropic must pay AWS for the cloud resources and then share 50% of the remaining gross profit with Amazon. According to Anthropic's business forecast from late last year, its compute costs paid to cloud providers, including Amazon, could reach $1.9 billion by 2026.
Internal Cost-Saving Measures
A former Amazon employee revealed that the company has long been concerned about potential future price increases from Anthropic and has proactively implemented several cost-reduction measures. Jigar Thakkar, Vice President of Enterprise AI Agents at AWS, cited the workplace assistant Quick as an example, which features a model selection function that automatically chooses the most cost-effective model from Bedrock's offerings for a given task.
Thakkar stated in an interview, "We are committed to matching the most suitable model for different business scenarios, with both cost and response speed being core considerations."
Intellectual Property Dynamics
Anthropic's heavy reliance on Amazon for compute deployment and sales channels has led to friction. Previously, the Bedrock development team was slow to integrate new Anthropic features onto the platform, causing dissatisfaction. Conversely, Amazon's technical teams faced numerous restrictions when attempting to customize or further develop Anthropic's models.
The persistent performance gap between Amazon's own models and Anthropic's has also caused concern within AWS management. The fact that many of Amazon's products are built on Claude technology risks creating a perception that Amazon's AI applications are merely wrappers around Anthropic's core technology.
Sources familiar with the matter said that anticipating potential future model price hikes, Amazon's technical teams have begun work on model distillation, creating lighter internal models based on Claude. Other business needs are also driving this project. Under their agreement, Amazon has the right to distill smaller parameter models from Anthropic's technology for internal business use.
The two companies have also collaborated on intellectual property initiatives. For instance, Anthropic participated in optimizing the software for the Trainium chip, significantly lowering the technical barrier for developers using this custom AI chip.
Comments