Claude 3.7 and new Cursor: first impressions

Anthropic finally released Claude 3.7, and Cursor has a new version with some interesting features. Let’s take a look at what’s new and some first impressions.

What’s new?

Claude 3.7 is Anthropic’s latest model release, bringing significant improvements across several dimensions:

  • Improved reasoning: Better at complex problem-solving, especially in mathematics, coding, and step-by-step reasoning tasks
  • Enhanced knowledge: More accurate and up-to-date information across domains
  • Reduced hallucinations: Less likely to make up information when it doesn’t know something
  • Longer context window: Now supports up to 200,000 tokens (approximately 150 pages of text)
  • Faster responses: Generates responses up to 2x faster than Claude 3.5 Sonnet
  • Better tool use: More capable at using tools like code interpreters and retrievers
  • Improved multilingual capabilities: Better performance across non-English languages

Cursor’s Latest Update

Cursor has also released a significant update with several new features that enhance the coding experience:

  • Agent mode by default: Agent is now the primary interface, creating a more streamlined and powerful AI experience. Gone is the confusion between different modes - you now get one intelligent interface that understands what you need.
  • Refreshed interface: They’ve started rolling out cleaner, more focus-oriented themes. The @-context menu has been simplified to help Agent deliver more precise results.
  • Integrated web search: Agent now searches the web automatically when needed, without you having to explicitly tell it to do so with @Web commands.
  • Better file exclusions: The .cursorignore file now does more - it prevents files from appearing in chat, blocks them from tab completions, and excludes them from indexing. For finer control, you can use the new .cursorindexingignore specifically for indexing rules.
  • Continuing past tool limits: When you hit the 25 tool call limit, you can now simply click “continue” to keep going (this counts as starting a fresh request).
  • Global project rules: You can now apply rules across your entire project, with helpful visual indicators showing when rules are active.
  • Enhanced MCP functionality: Agent can automatically run MCP tools in Yolo mode. You can configure Project Servers using /.cursor/mcp.json, and Agent can now leverage MCP resources as context.

First impressions

Off the top: it’s really good. I’ve been using it for a day or two now, so not too long, but Claude 3.5 was already fantastic for my normal python DS/web development and 3.7 is certainly the same or better in terms of quality of responses.

The most stark difference in actual usage has been the speed. Between the cursor updates and the model model updates it’s shocking. In the past I did not value the speed of the models that much honestly. I was using mostly agent mode, and with large contexts and many chained calls, it would be a few minutes of cursor churning through a big problem before I really had something to review or do. Because of that, I’d often get distracted, or have this really running in the background while I did something else. Now, the responses are so fast that I can actually get into a flow state of development while using 3.7 and agent mode. So while I’m not sure the model is really that much faster, the total time of development is dramatically faster because I’m engaged the whole time and not just wandering by the computer while doing chores to check in.

In the new cursor update, I’ll also note that the always on web use and faster file searches seem to have made the agent loops much faster and more more reliable.

Overall, I’m pretty happy with it, but will keep using it and report back here as I get more comfortable with it.