Amazon Could Pay More for Anthropic Technology Under New Deal
The Takeaway
- Anthropic renegotiated deal with Amazon to one based on tokens rather than compute hours
- The new arrangement could make it more expensive for Amazon to use Anthropic models.
- Amazon is evaluating other AI models to mitigate rising Anthropic costs.
As Anthropic grows to become a leading AI model used by enterprises, it’s been flexing its power with customers—including one of its most important early backers, Amazon.
Earlier this year, Anthropic renegotiated an aspect of its partnership with Amazon, making it more expensive for the cloud and e-commerce giant to use Anthropic models in its products, according to two people with knowledge of the discussions between the two firms. As a result, Amazon is evaluating whether it can save on costs by using other models, including OpenAI’s and Amazon’s own Nova models, one of the people said.
Under the new pricing arrangement, to take effect next year, Amazon will pay for Anthropic models based on the number of tokens, the bits of information processed by an AI model, according to two people with knowledge of the conversations. Previously, Amazon paid based on the number of computing hours it used, the people said.
The change could significantly increase Amazon’s costs. Amazon uses Anthropic models to power a bevy of AI products it offers to businesses and consumers, including shopping assistant Alexa for Shopping, coding tool Kiro and workplace assistant Quick.
“Amazon and Anthropic share a multifaceted partnership grounded in technical collaboration, and we continue to foster that relationship and deepen our work together,” an Amazon spokesperson said. “It’s incorrect that changes from our expanded collaboration will increase our costs.”
Despite recent changes to how Anthropic bills Amazon and other customers, Anthropic said the overall cost of the models has been going down.
“The cost of getting important work done with Claude falls every generation,” an Anthropic spokesperson said. “In November 2025 we significantly reduced Opus pricing, and that price has held since while the models keep getting more capable, so the same budget buys materially more each cycle.”
The pricing change comes as Amazon is getting closer to Anthropic’s archrival, OpenAI. Amazon agreed to invest up to $50 billion in the company earlier this year. As part of that deal, OpenAI will use Amazon Web Services infrastructure and AWS will sell OpenAI’s models, plus Amazon gets access to OpenAI’s tech to use inside its products. Amazon also agreed to invest up to $25 billion more in Anthropic earlier this year.
More recently, Amazon has played a complicating role in Anthropic’s simmering conflict with the U.S. government. CEO Andy Jassy raised concerns with White House officials about security risks in Anthropic’s latest models. That call contributed to the White House prohibiting foreign nationals from using those models, called Mythos and Fable, which subsequently led Anthropic to suspend use of the models more broadly.
An Anthropic spokesperson said that their relationship with Amazon continues to grow. “Amazon is one of our most important partners, with more than 100,000 customers building with Claude on AWS,” a spokesperson said.
Tokens vs. Hours
Tokens are the standard measure of charging for models. Many Anthropic customers have been hit with steeper costs this year as Anthropic switched its pricing model from seat-based pricing that capped at $200 a month for some plans to usage-based pricing. Some customers have been eating those costs, while others have responded by adopting open-source models. (Amazon’s arrangement had been structured as a usage-based model, linked to compute hours, already).
But Amazon is one of Anthropic’s most important business partners. Their business arrangement dates back to 2023, when Amazon agreed to invest $4 billion in Anthropic, in exchange for Anthropic making AWS its primary cloud provider and using Amazon’s custom AI chips, Trainium and Inferentia.
AWS provided Anthropic with the computing power to build its models, along with a huge swath of new business customers. AWS debuted a new service in 2023: Bedrock, where its customers could access an array of AI models.
Anthropic had to pay handsomely for that relationship. When AWS sells Anthropic models, Anthropic not only pays AWS to use its cloud service but forks over half of what is left (its gross profits) to Amazon, The Information previously reported. Anthropic expected to pay cloud providers including Amazon $1.9 billion in 2026, according to the company’s projections from late last year.
Amazon employees have been worried for some time about the risk of Anthropic raising prices, according to a former employee.
The company has taken various measures to keep costs down for customers. Jiggar Thakkar, vice president of agentic AI for business at AWS, said its workplace assistant Quick for instance, allows customers to choose what kind of AI model they want for their task and then Amazon selects the model from the many models available on Bedrock.
“We picked the right model for the right task. And yes, cost is a consideration, speed is a consideration,” Thakkar said in an interview.
Intellectual Property
Anthropic’s reliance on Amazon to both host and sell its models has led to tensions. Anthropic was frustrated when Bedrock engineers didn’t add new Anthropic features fast enough to the Bedrock service. Amazon engineers, for their part, have been limited in how much they can customize Anthropic models.
Amazon struggled to create its own large language models that performed as well as Anthropic’s. AWS executives worry that too many Amazon products are powered by Anthropic rather than by Amazon’s own Nova models, which could lead to the perception that its products are just a wrapper on top of Anthropic’s technology.
Fears that Anthropic’s models might eventually become more expensive have prompted some engineers to distill them proactively, according to someone involved with the effort. Employees have also distilled models for other reasons, the person said.
Amazon has some rights to use Anthropic models to build small models for internal use cases, a source familiar with the arrangement said.
There have been times when Anthropic and Amazon have taken a collaborative approach to sharing intellectual property. Those include contributions Anthropic made to Trainium’s software, which helped make the chip easier for developers to use.