Anthropic PBC today launched Claude 3.5 Sonnet, the company’s first release in a forthcoming artificial intelligence large language model family that outperforms both competing models and its Claude 3 Opus model, which was introduced three months ago.

The AI research startup said the new model comes out of the box operating at twice the speed of Claude 3 Opus, which makes it ideal for complex tasks such as context-sensitive customer support and automating multi-step workflows.

According to the company, the new Sonnet evaluated higher than industry-leading LLMs on numerous benchmarks, setting new industry levels for intelligence for graduate-level reasoning or GPQA, undergraduate-level knowledge or MMLU, and coding proficiency, known as HumanEval. Claude 3.5 Sonnet hit 92.0% in zero-shot code evaluation compared with 90.2% for GPT-4o, OpenAI’s flagship model, and 84.1% for Google LLC’s Gemini 1.5 Pro.

The startup said that when provided with the proper tools, the model can independently write, edit and execute code with highly sophisticated reasoning capabilities exceeding Claude 3.0. Applied to code translation, this can make developer migration among different codebases or updating legacy applications to new frameworks much easier.

The model also contains the company’s most powerful vision model yet built for visual reasoning, providing it with key capabilities for understanding and interpreting written language and symbols. Users can write down complex math problems, charts and graphs and the model can quickly ingest them and use them for math problem-solving or other reasoning tasks.

Claude 3.5 Sonnet can handle text transcription even from imperfect images. That’s a problem for retail, logistics and financial services, where it’s not always easy to get a clear picture of a transcript, notepad or a receipt.

“Claude Sonnet 3.5 has significantly enhanced our Legal AI Assistant’s capabilities,” said Richard Robinson, chief executive of Robin AI Ltd., the developer of an AI “copilot” for reviewing and producing legal documents. “Claude Sonnet 3.5 outperformed Opus or GPT4o in our testing and is exactly the sort of leap forward we need to keep demonstrating value to our customers. We’re seeing improvements in the speed and accuracy of contract reviews, as well as more nuanced interpretations of complex legal language. This is allowing legal professionals to focus more on strategic work while the AI handles routine document analysis.”

As part of this release, Anthropic is also introducing Artifacts, a new feature for Claude.ai, the web interface. It adds a dedicated window alongside the chat box next to the chatbot. When a user asks Claude to generate code, text documents, or website designs, this region will produce a workspace where users can see, edit and build upon a real-time representation of what they’ve requested.

The Artifacts workspace, now in preview, is designed to provide a real-time vision for collaborative work alongside Claude. Before, users would have to ask for a code snippet or website design and copy their code into an editor or other interface to view its end result. Now they can see it immediately and iterate on it directly within their chat with the LLM.

The new model is now available for free to access through Claude.ai and the Claude iOS app. Claude Pro and Team plan subscribers can access it with higher rate limits. The LLM is also available through the Anthropic application programming interface and on managed services such as Amazon Bedrock and Google Cloud’s Vertex AI.

The model costs $3 per million input tokens and $15 per million output tokens and it supports prompts up to 200,000 tokens, units of data that each hold a few letters or numbers. This is the same prompt context size as the Claude 3 family of models.

With the release of Claude 3.5 Sonnet, Anthropic seeks to improve the tradeoff curve among intelligence, speed and cost. Over the next few months, the startup said, it will release the rest of the Claude family including the lightweight Claude 3.5 Haiku and the larger Claude 3.5 Opus later this year.

Image: Anthropic