Amazon, the leading investor in artificial intelligence startup Anthropic, is relying on that company’s AI models to power the most advanced capabilities in new Alexa devices, according to two people familiar with the matter.
Anthropic’s Claude large language model is handling the vast majority of questions that customers have been asking the new Alexa, according to the people, who asked not to be named because the details are confidential.
Earlier this week, Amazon unveiled its long-awaited overhaul of its decade-old devices. For the first time, Amazon will charge users to access a version of Alexa. The “Alexa+” service will be $19.99 a month, or free for Amazon Prime members, and will roll out in early access next month.
During the demonstration this week, Alexa+ was able to make dinner reservations, order groceries or book an Uber, tasks that were mostly not possible with older versions. Alexa, once a leader in natural language and machine learning, has been lagging the competition since the emergence of generative AI chatbots such as OpenAI’s ChatGPT, which have quickly advanced beyond text conversations and into AI-generated audio, images and videos.
Anthropic declined to comment. Amazon said information in the story is “false.”
“In fact, over the past four weeks Nova handled over 70% of conversations — including complex requests,” an Amazon spokesperson wrote in an email. “That said, from a customer perspective this doesn’t matter — both are excellent models and in there to deliver the best experience for customers.”
The spokesperson added that the way Amazon designed the architecture, “Alexa+ is always using the best model for each task.”
At the event, Amazon CEO Andy Jassy described the overhaul as “rearchitecting” the brain of Alexa.
In addition to the roughly $8 billion its agreed to invest in Anthropic, Amazon has been developing its own AI models and rolled out its Nova series late last year. Amazon offers Anthropic’s Claude model as part of its Amazon Web Services Bedrock offering, which gives customers access to a variety of AI options. It lets users pick between Amazon’s Nova and Titan models, as well as Mistral and others.
Amazon said it used Bedrock to power Alexa. But Claude has been the model tackling the more complex tasks the company showcased at its devices event in New York this week, the sources said. Claude has powered the queries that take more thinking and “intellectual heft,” one person said.
While Amazon’s own AI models are still being used, they are handling the tasks that require less reasoning, these people said.
As part of Amazon’s initial investment into Anthropic, the company was able to use a certain amount of Anthropic capacity for free over the course of 18 months, according to one source. That deal has now expired, and the companies are renegotiating the terms of their agreement, the person said.
Anthropic’s model has expanded far beyond Alexa within Amazon and is helping to drive product search and advertising as well, the person said.
Panos Panay, Amazon’s senior vice president of devices and services and the leader of Alexa’s redesign, called Anthropic an “awesome” partner at this week’s event. Panay, who joined the company in 2023 after two decades at Microsoft, said Anthropic’s foundational model is “incredible.”
“We pick the model that’s right for the job,” Panay told CNBC in an interview on Wednesday. “We use Amazon Bedrock — Alexa picks the right model to get the job done.”
