Anthropic Releases New AI Model Claude 3.5 Sonnet

Anthropic unveiled their newest AI model, Claude 3.5 Sonnet, on Thursday, touting it as their most intelligent model to date. This upgraded version is said to be twice as fast as its predecessor, Claude 3 Opus, and significantly more cost-effective, with a five-fold reduction in operational expenses. Following in the footsteps of their competitor, OpenAI, who recently released ChatGPT’s latest iteration, GPT-4o, Claude 3.5 Sonnet is now freely accessible to all web and iOS users, as well as developers.

According to Michael Gerstenhaber, a product manager at Anthropic, Claude 3.5 Sonnet now holds the title of the world’s most intelligent model. “We are witnessing the early stages of a rapid expansion in this industry,” Gerstenhaber commented in an interview with TIME prior to the launch.

The rapid pace of model releases further emphasizes this point. Claude 3.5 Sonnet arrives merely three months after Anthropic’s previous suite of models, Claude 3, and less than a year after the introduction of its predecessor, Claude 2.
Unlike GPT-40, the latest version of Claude does not have the ability to search the internet or generate image files. Instead, its “intelligence” is evaluated based on its performance on various benchmarks, which, although not perfect, indicate that it is currently leading the way. However, what truly excites Gerstenhaber, the creator of Claude, are the qualitative aspects of the model. Early users of the 3.5 Sonnet version have praised its intelligence and the humor it exhibits during interactions. This can be attributed to Anthropic’s deliberate efforts to give Claude a unique personality and to foster a genuine curiosity about the perspectives and values of the individuals it engages with, as described on the company’s website.
In addition to its enhanced coding and image transcription capabilities, version 3.5 of Sonnet introduces a new feature called “artifacts,” which revolutionizes the functionality of chatbots. Anthropic, the company behind Sonnet, describes artifacts as a significant development in the way chatbots operate. When users request tasks such as drafting documents, generating code, or assisting with web design, Sonnet presents the content in a separate panel alongside the chat interface. Users can observe real-time updates as they make changes to the content.

The introduction of artifacts provides insight into Anthropic’s vision for the future of its models. It suggests that these models will transcend the boundaries of traditional chatbots and serve as collaborative platforms, enabling teams to work together seamlessly. In this future scenario, Sonnet, acting as an on-demand teammate, will facilitate collaboration on various projects. This concept aligns more closely with collaborative software tools like Notion and Google Docs than with the chatbot interfaces that are prevalent today.
Anthropic affirms that Sonnet 3.5 underwent rigorous safety testing. Concerns regarding the capabilities of current and future models are prevalent among major AI laboratories, many of which have published documents outlining their plans based on the perceived danger of their models, commonly referred to as “responsible scaling policies” (RSPs). Despite its increased intelligence, Sonnet 3.5 is still classified as ‘AI Safety Level 2’ according to Anthropic’s RSP, the same level Claude 3 Opus was at in March.

Prior to its release, Sonnet 3.5 underwent pre-deployment testing in collaboration with the United Kingdom’s AI Safety Institute. The results of this testing were also shared with the AI Safety Institute in the United States, as part of a partnership between the two institutions.
More information on the safety performance of Sonnet 3.5 in various tests reveals that the model does not pose any significant risks, such as those related to bioweapons or nuclear war. However, it is noted in the documentation that Anthropic observed an increase in capabilities in areas relevant to risk compared to Claude 3 Opus.

In terms of privacy, Anthropic states that they do not train their generative models using user-submitted data unless explicit permission is given by the user. They also mention that updated versions of Opus and Haiku, which are respectively larger and smaller versions of Sonnet, will be released later this year to complete the Claude 3.5 model family.

Additional reporting by Billy Perrigo from London.

For further inquiries, please contact us at [email protected].