Amazon.com Inc. will reportedly use technology from Anthropic PBC to power an upcoming paid edition of Alexa.

Reuters today cited sources as saying that the subscription-based voice assistant could make its debut in October. Amazon is expected to preview Remarkable Alexa, as the offering will reportedly be called, at a product event in September.

San Francisco-based Anthropic is one of OpenAI’s main competitors in the large language model market. It has raised more than $7 billion to date, most of which was provided by Amazon through an investment that closed in March.

The paid edition of Alexa is expected to include several features not available in the current version. The upgraded assistant will reportedly be capable of taking into account users’ past questions, as well as its own answers to those questions, when responding to prompts. Remarkable Alexa will reportedly use that capability to aggregate news articles, generate shopping suggestions, provide fashion tips and perform a range of other tasks.

The service is reportedly capable of not only retrieving information but also carrying out actions on consumers’ behalf. Ordering food and drafting emails are two of the supported use cases listed in today’s report. Moreover, it appears that the service can perform some of those tasks with a certain degree of autonomy: Remarkable Alexa can reportedly set a morning alarm even if the user forgets to do so.

Amazon is expected to integrate the service into its smart home product portfolio. According to Reuters, the upgraded Alexa will be capable of interacting with connected devices such as TVs.

Amazon will reportedly rely primarily on Anthropic’s Claude series of LLMs to power Remarkable Alexa. The newest edition to the model family, Claude 3.5 Sonnet, debuted in June. It managed to outperform OpenAI’s flagship GPT-4 across seven of the nine task collections that Anthropic evaluated in an internal benchmark test.

Claude 3.5 Sonnet is the first model in a planned suite of three LLMs. It’s set to be joined by a more capable model called Claude 3.5 Opus, as well as a scaled-down version known as Claude 3.5 Haiku. The latter LLM will trade off some output accuracy for increased hardware efficiency and response speed.

Today’s report suggests that the paid version of Alexa will use not just one but multiple models from the Claude series. It’s possible Amazon plans to have a high-end model, such as Opus, process the most complicated user prompts and route simpler requests to a smaller Claude LLM such as Haiku. Such an arrangement could help reduce inference costs and thereby improve Remarkable Alexa’s profit margins. 

Amazon is expected to charge $5 to $10 per month for the service. According to a Bank of America analysis cited by Reuters, Alexa currently has 100 million active users and about one in ten might be willing to pay for a more capable version. That means Amazon could generate $1.2 billion in annual revenue from the paid version of the assistant if it charges $10 per month.

OpenAI recently started training a successor to GPT-4o. If that next-generation LLM proves to be more capable or cost-efficient than Claude, the likelihood of an Alexa integration might increase. Today’s report suggests that Amazon could delay or alter its plans to integrate Claude into the AI assistant if certain internal benchmarks are not met. 

Photo: Amazon