The AI company Anthropic, founded by several former OpenAI employees, claims that the performance of the new Claude 3 series AI model is comparable to or better than leading models from Google and OpenAI. Unlike earlier versions, Claude 3 also has multimodal capabilities, allowing it to understand both text and image inputs.

Anthropic states that Claude 3 will be able to answer more questions, understand longer instructions, and be more accurate. Claude 3 can comprehend more context, which means it can handle more information. Claude 3 has three models: Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus, with Opus being the largest and "most intelligent" model. Anthropic states that Opus and Sonnet are now available on claude.ai and its API. Haiku will be released soon. All three models can be deployed on chatbots, autocomplete, and data extraction tasks.

Earlier versions of Claude refused to answer some harmless prompts, which the company wrote "indicated a lack of contextual understanding." The new model is less likely to refuse to answer prompts that touch on safety concerns, similar to rumors about Meta's plans when releasing Llama 3.

Anthropic claims that the Claude 3 model can provide near-instantaneous results even when parsing dense materials like research papers. A blog post states that the minimal version of Claude 3, Haiku, is the "fastest and most cost-effective model on the market," capable of reading a research paper with charts and graphs in "less than three seconds."

Anthropic states that Opus outperforms most models in several benchmark tests. It scored 50.4% in graduate-level reasoning tests, surpassing OpenAI's GPT-4, which scored 35.7%. It also performs better in answering math questions, writing code, and understanding reasoning.

The new model also shows significant improvements compared to the previous Claude 2.1 model. The medium-sized Sonnet model is twice as fast as Claude 2 and Claude 2.1. "It excels in tasks that require quick responses, such as knowledge retrieval or sales automation," according to Anthropic.

Anthropic trained the Claude 3 model using non-public internal and third-party datasets, as well as publicly available data, before August 2023. The company states in the paper introducing these three models that they were trained using hardware from Amazon AWS and Google Cloud. Both companies have invested in Anthropic, with Amazon investing $4 billion. Claude 3 will be available in AWS's model library Bedrock and Google's Vertex AI.