Anthropic Touts New AI Model as ‘Most Intelligent Yet’

Anthropic launched a new AI model Thursday which it says is its “most intelligent model yet.”

The new model, Claude 3.5 Sonnet, is reportedly twice as fast as Claude 3 Opus, the company’s previous best-in-class AI, and five times cheaper to run. Following a trend set by its competitor OpenAI—which just last month released the newest version of ChatGPT, GPT-4o—Claude 3.5 Sonnet is free for all users on the web and iOS to access. It has also been made available to developers.

[time-brightcove not-tgx=”true”]

Claude 3.5 Sonnet is “now the most intelligent model in the world,” claims Michael Gerstenhaber, a product manager at the company. “We are at the beginning of a Cambrian explosion of this industry,” Gerstenhaber told TIME ahead of the release.

Looking at the pace of model releases, it’s easy to see why. Claude 3.5 Sonnet comes just three months after the release of Anthropic’s previous suite of models, Claude 3, and just under a year after the release of its ancestor, Claude 2.

Unlike GPT-4o, the latest version of Claude cannot search the internet or generate image files. Its “intelligence” is measured by its performance on a series of benchmarks, which, while imperfect, do show that it is currently at the front of the pack. But it’s the qualitative aspects of the model, rather than its performance on benchmarks, that Gerstenhaber is most excited about. Early users of 3.5 Sonnet singled out its intelligence and the humor it displayed in interactions, he said. This is likely the result of Anthropic’s efforts to give Claude a distinctive character, and to have the model “display a genuine curiosity about the views and values of the people it’s talking with,” according to the company’s website.

Read More: Inside Anthropic, the AI Company Betting That Safety Can Be a Winning Strategy

In addition to its improved coding and image transcription abilities, 3.5 Sonnet has a new feature that changes how chatbots work—what Anthropic calls “artifacts.” When users ask Claude to draft documents, generate code, or assist with something like web design, it displays the requested content in a separate panel that sits next to the chat. Ask it to make changes to the content, and you can see it update in real-time.

Artifacts offer a glimpse into how Anthropic is thinking about the future of its models, where they go beyond chatbots and act more as sandboxes where teams of people can collaborate on work, “with Claude serving as an on-demand teammate,” according to the company’s press release. This suggests that future models may have more in common with collaborative software tools like Notion and Google Docs than the chatbot interfaces proliferating today.

Anthropic says Sonnet 3.5 was subject to rigorous safety testing. Concerns about the capabilities of current and future models are on the minds of virtually all major AI labs, most of which have published documents detailing what actions they will take based on how dangerous they deem their models to be—often called “responsible scaling policies” (RSPs). Despite the increase in intelligence, the model is still considered to be at ‘AI Safety Level 2’ according to Anthropic’s RSP; the same level that Claude 3 Opus was at in March.

Sonnet 3.5 was shared with the United Kingdom’s AI Safety Institute for pre-deployment testing ahead of its release. The results of this testing were also shared with the United States’ AI Safety Institute, as part of a partnership between the two institutions.

Further information about how Sonnet 3.5 performed under safety tests reveals that the model “does not pose risk of catastrophic harm,” for example by increasing risks related to bioweapons or nuclear war. However, the documentation does note that Anthropic “observed an increase in capabilities in risk-relevant areas compared to Claude 3 Opus.”

Regarding privacy, Anthropic says they “do not train [their] generative models on user-submitted data unless a user gives us explicit permission to do so.” They also note that “to complete the Claude 3.5 model family,” updated versions of Opus and Haiku (respectively bigger and smaller versions of Sonnet), will be out later this year. 

-With additional reporting by Billy Perrigo/London