Discover Claude 3.5 Sonnet & Haiku with New Computer Use Capabilities
Explore Claude 3.5 Sonnet and Haiku, featuring advanced coding skills and innovative computer use for seamless automation.
We’re excited to introduce two major updates to our AI family: the upgraded Claude 3.5 Sonnet and a brand-new model, Claude 3.5 Haiku. These models come packed with powerful new features for developers, especially those working on complex coding tasks and UI automation. Plus, we’re rolling out something truly groundbreaking — a feature called computer use, now in public beta. It lets Claude interact with computer screens like a human, navigating interfaces, typing, and clicking through tasks. Let’s dive into what these new tools bring to the table and what they mean for the future of AI.
Claude 3.5 Sonnet: The New Coding Powerhouse
The Claude 3.5 Sonnet update pushes the envelope on coding, especially when it comes to complicated, multi-step tasks. Here’s what’s new:
- Enhanced Coding Skills: On key benchmarks like SWE-bench Verified, Claude 3.5 Sonnet jumped from 33.4% to 49% accuracy, putting it ahead of every other publicly available AI model in coding, including specialized systems.
- Expanded Tool Use: Whether you’re working in retail or airlines, Claude 3.5 Sonnet handles complex tool use tasks better than before, scoring 69.2% on the TAU benchin the retail sector and 46% in the airline sector — both substantial improvements.
- Real-World Results: Companies testing it, like GitLab, found a 10% boost in reasoning accuracy with no added latency, making it a go-to choice for demanding tasks. The Browser Company also used Claude 3.5 Sonnet for automating web workflows and said it outperformed any other model they’d tried.
We put Claude 3.5 Sonnet through a rigorous testing process with industry safety experts, ensuring it meets standards for responsible scaling. So whether you’re building new apps or automating tasks, Sonnet is ready to deliver.
Claude 3.5 Haiku: Speed and Precision, Redefined
If speed and affordability are what you need, Claude 3.5 Haiku was built for you. This model takes the fast, efficient performance of Claude 3 Haiku and kicks it up a notch.
- Power-Packed Yet Efficient: Scoring 40.6% on the SWE-bench Verified, Haiku outpaces both the original Claude 3.5 Sonnet and many of today’s leading models.
- Built for Real-Time Interaction: With low latency and precise instruction-following, Claude 3.5 Haiku is ideal for user-facing applications or processing massive datasets — think customer insights, pricing analysis, and more.
Claude 3.5 Haiku will roll out later this month across our API, Amazon Bedrock, and Google Cloud’s Vertex AI, with text-only support at first and image input capabilities on the way.
A New Way to Work: Teaching Claude to Use Computers Like a Human
We’re also launching something fundamentally new: a computer-use feature that lets Claude interact with computers as a human would. In this public beta, developers can teach Claude to look at screens, move cursors, click buttons, and even type out information. It’s a leap in AI usability, with exciting potential for automating complex workflows.
- More than Just a Tool: Claude can take on intricate tasks like data entry or research across multiple web pages. Companies like Replit are using it to evaluate apps as they’re built, and others are exploring its potential to save time on repetitive processes.
- Performance on OSWorld: Claude scored 14.9% in OSWorld’s “screenshot-only” category and 22.0% with added steps — outperforming any other AI model in computer navigation tasks.
- Safety First: While computer use is promising, we’re cautious about safe deployment. We’ve added classifiers to detect potentially harmful activities and ensure ethical use, especially in areas prone to issues like spam or fraud.
This beta is still experimental, so some actions like scrolling or zooming may not be flawless yet. We’re encouraging developers to explore this with low-risk tasks and look forward to their feedback to refine and expand these capabilities.
Looking Forward
Claude 3.5 Sonnet and Claude 3.5 Haiku bring us closer to a future where AI can intuitively assist with real-world tasks, whether it’s coding complex systems or navigating screens like a human. And with our public beta for computer use, we’re just scratching the surface of what’s possible. Developers at Asana, Canva, DoorDash, and beyond are already pushing these boundaries, and we can’t wait to see what they create.
Both models are now available on the Anthropic API, Amazon Bedrock, and Google Cloud’s Vertex AI. Give them a try, explore the possibilities of computer use, and join us in shaping the next frontier of AI!
