Anthropic Releases Claude 3

This week saw the release of Claude 3 by AI startup Anthropic, which it claims set new industry benchmarks with “near-human” capability that beats OpenAI’s GPT-4 in performance.

Claude 3 is a family of models, namely Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. Opus is the most powerful, while Haiku is the most “compact” and offers the fastest performance.

Better than GPT-4

Claude 3 is also the first multimodal AI from Anthropic, and unlike other popular models can analyze up to 20 images in a single request. This allows it to compare and contrast images, though it cannot produce images.

Claude 3 will initially support a 200,000-token context window, equivalent to about 150,000 words. It is capable of supporting up to 1-a million-token context window, however, though this is currently only offered to select customers.

According to Anthropic, when compared to Claude 2.1, Opus demonstrates a twofold improvement in accuracy (or correct answers) on these challenging open-ended questions while also exhibiting reduced levels of incorrect answers.

Claude 3 Opus beats GPT-4 on 10 AI benchmarks. As noted by Ars Technica, however, several of the wins are very narrow. Others are larger, though what it means to users is much more difficult to pin down. Still, it is worth noting that no other model has beaten GPT-4 on such a range of benchmarks like this.

Room for further improvement

“We do not believe that model intelligence is anywhere near its limits, and we plan to release frequent updates to the Claude 3 model family over the next few months,” notes Anthropic in a blog entry introducing Claude 3.

The firm promised new features such as “tool use” or the ability to call functions, interactive coding in the form of REPL, or Read-Eval-Print Loop, and more advanced agentic capabilities. Crucially, the models will soon cite the source of their answers to questions for users to verify.

Claude 3 is hardly perfect, of course. Unlike some generative AI rivals, Claude 3 can’t search the web. Claude 3 can only answer questions using data from before August 2023.

Claude 3 is available now, with the mid-tier Sonnet accessible for free while the Opus is subscription-only and priced similarly to ChatGPT Plus at USD20 per month. Haiku will only be available later this year.

You can access the technical white paper about Claude 3 here (pdf).

Image credit: iStockphoto/Adrian Vidal