Experience the frontier model for codingExperience the frontier model for coding

There's a reason Claude is the default model for all Cursor users—Anthropic's approach to building models delivers on real-world tasks. During our extensive testing of Claude 3.7 Sonnet, we've seen significant improvements in the model's ability to understand and handle complex codebases and multi-step tasks. Now with two ways to think, Claude cements its place as the industry leader for coding.

Claude 3.7 Sonnet is absolutely epic for coding—it delivers complete production-grade code with genuine design taste while maintaining context through complex iterations. This will enable us to empower everyone to add interactivity and smarts to their Canva designs, even without coding skills.

What impresses us most about Claude 3.7 Sonnet is its instruction following and multi-turn tool calling accuracy. Coupled with its extended output window, 3.7 Sonnet is transformative for code generation and agentic workflows.

Testing Claude 3.7 Sonnet's abilities building web apps zero-to-one reveals advancements far beyond other models. From consumer sites to dashboards, it's delivering sophisticated frontends with remarkable design quality—creating powerful applications that instantly demonstrate their value.

What we're seeing with Claude 3.7 Sonnet in our development environment is very impressive. From multi-file edits to enhanced reasoning output, it's bringing unprecedented capabilities to complex coding workflows for Bolt.new.

Claude 3.7 Sonnet represents a leap in computer use—its ability to determine correct agentic actions is significantly improved over GPT-4o and prior Claude versions. Complex actions show far lower error rates with less guidance needed, reducing 4-hour automated testing to just 10 minutes.

Claude 3.7 Sonnet has transformed AI-powered software engineering. At iGent AI, we've measured 7x success on complex tasks and 2.5x fewer context errors. Engineers now build in days what took weeks, with a thought partner that preserves creative control.

Claude 3.7 Sonnet is hands down the best model we've used for programming, especially Deno. It creates complex applications and handles extensive refactors in a single turn, setting a new bar for AI assistance across our 30,000+ users across the enterprise!

What impresses us about Claude 3.7 Sonnet is how it combines rapid analysis with nuanced code understanding. When planning code updates, it considers subtle context such as files being generated from a template. It also filters out 21% more noisy security alerts than our previous model.

What stands out about Claude 3.7 Sonnet is its natural and human-like approach to organizing information. From thoughtful code documentation to well-structured summaries, it consistently focuses on what matters most.

Per our internal testing results, we're seeing delightful improvements using Claude 3.7 Sonnet for web story generation—it understands narrative context deeply and delivers more natural dialogue than other models we've tried. We are not just seeing richer and longer outputs, but more enhanced emotional intelligence and cultural adaptability.

Results from our early testing of Claude 3.7 Sonnet in GitHub Spark demonstrate two large improvements from its predecessor: it generates higher quality apps (e.g. feature set, user interface) from a brief natural language description, and in thinking mode it is more successful at generating passing code across iterations.

Claude 3.7 Sonnet has been very impressive at planning code changes—we've seen cases where it's far better than any other model. Even with thinking off, it effectively handles medium-size features requiring both frontend and backend updates.

We're excited to continue our partnership with Anthropic in bringing Claude 3.7 Sonnet to Poe, which combines deep reasoning and performance with engaging conversation skills. 3.7 Sonnet will be especially helpful for our users who want to solve real-world coding tasks.

Claude 3.7 Sonnet is particularly impressive in codebase understanding. We saw 10+% gains on agentic QA, essentially solving our benchmarks even without reasoning mode enabled—a change on par with the step from Haiku to Sonnet.