Today Anthropic launched Claude 3.5 Sonnet — the first release in the upcoming Claude 3.5 model family. Claude 3.5 Sonnet raises the bar for industry intelligence, outperforming competing models and Claude 3 Opus across a wide range of evaluations, while operating at the speed and cost of Claude 3 Sonnet's mid-tier model.
We at GPTunneL have already integrated the Sonnet 3.5 model, and it's available to everyone on our website, as well as in our Telegram bot!
Advanced intelligence at 2x the speed
Claude 3.5 Sonnet sets new industry benchmarks for graduate-level reasoning (GPQA), undergraduate-level knowledge (MMLU), and coding proficiency (HumanEval). It shows a marked improvement in understanding nuance, humor, and complex instructions, and it excels at writing high-quality content with a natural, relatable tone.
The model runs twice as fast as Claude 3 Opus. This performance boost, combined with efficient pricing, makes Claude 3.5 Sonnet ideal for complex tasks such as context-sensitive customer support and orchestrating multi-step workflows.
In an internal evaluation Claude 3.5 Sonnet solved 64% of problems, outperforming Claude 3 Opus, which solved 38%. The evaluation tests a model's ability to fix bugs or add functionality to open-source code based on a natural-language description of the desired improvement. Given instructions and the appropriate tools, Claude 3.5 Sonnet can independently write, edit, and run code with advanced reasoning and troubleshooting capabilities. It handles code translation with ease, making it especially effective for modernizing legacy applications and migrating codebases.
Advanced vision
Claude 3.5 Sonnet is the strongest vision model to date, outperforming Claude 3 Opus on standard vision benchmarks. These quality improvements are most noticeable in tasks that require visual reasoning, such as interpreting charts and graphs. Claude 3.5 Sonnet can also accurately transcribe text from imperfect images — a key capability for retail, logistics, and financial services, where AI can extract more information from an image, chart, or illustration than from text.
Artifacts — a new way to use Claude
Today the company also introduced Artifacts on Claude.ai — a new feature that expands the ways users can interact with Claude. When a user asks Claude to generate content such as code snippets, text documents, or web designs, these Artifacts appear in a dedicated window alongside the conversation. This creates a dynamic workspace where users can see, edit, and iterate on Claude's creations in real time, seamlessly integrating AI-generated content into their projects and workflows.
This experimental feature marks Claude's evolution from a conversational AI into a collaborative work environment. It's just the beginning of a broader vision for Claude.ai, which will soon expand to support team collaboration. In the near future, teams — and eventually entire organizations — will be able to securely centralize their knowledge, documents, and ongoing work in one shared space, with Claude acting as an on-demand assistant.
Commitment to safety and privacy
The company's models undergo rigorous testing and are trained to reduce the potential for misuse. Despite the leap in intelligence in Claude 3.5 Sonnet, red-team evaluations showed that it remains at ASL-2.
As part of its commitment to safety and transparency, the company brought in external experts to test and refine the safety mechanisms in this latest model. Claude 3.5 Sonnet was recently provided to the UK AI Safety Institute (UK AISI) for a pre-deployment safety evaluation. UK AISI tested 3.5 Sonnet and shared the results with the US AI Safety Institute (US AISI) under a Memorandum of Understanding made possible by the partnership between the US and UK institutes announced earlier this year.
The company incorporated policy feedback from external subject-matter experts to ensure the reliability of its evaluations and to account for emerging misuse trends. This collaboration helped the teams scale 3.5 Sonnet's evaluation capabilities across various types of misuse. For example, they used feedback from child-safety experts at Thorn to update classifiers and further fine-tune the models.
One of the core constitutional principles guiding the company's AI model development is privacy. They do not train their generative models on data submitted by users unless the user has explicitly permitted it. To date, the company has not used any customer or user data to train its generative models.
Coming soon
The company's goal is to meaningfully improve the trade-off between intelligence, speed, and cost every few months. To complete the Claude 3.5 model family, Claude 3.5 Haiku and Claude 3.5 Opus will be released later this year.
Beyond work on the next generation of the model family, the company is developing new modalities and features to support more business use cases, including integration with enterprise applications. The team is also exploring features like Memory, which will let Claude remember a user's preferences and interaction history within a specified scope, making the experience even more personalized and effective.
The company is continuously working to improve Claude and welcomes user feedback. You can send feedback about Claude 3.5 Sonnet directly within the product to help shape the development roadmap and improve your experience. As always, the company can't wait to see what you build, create, and discover with Claude.