Cursor offers a ton of different AI models from leading providers like OpenAI (GPT series), Anthropic (Claude series), and Google (Gemini), along with its own custom models and support for local models. Different models excel in different areas, and choosing wisely can significantly impact quality, latency, and cost.

Here’s a guide to model selection in Cursor, detailing when to use what model for optimal results:

Model Selection Strategy in Cursor

Claude Models (Claude 3.5 Sonnet, Claude 4 Sonnet, Claude 4 Opus)

• Strengths & Ideal Use Cases:
  • Frontend Development, UI/UX Design: Frequently praised for these tasks.
  • Code Simplification and Refactoring: Excellent at clarifying and restructuring complex code.
  • Complex Architecture and Critical Tasks: Claude 4 Opus offers top performance for these.
  • General Coding Tasks: Claude 3.5 Sonnet is often considered the “king of AI coding” and integrates best with coding tools, preferred by many advanced users for most coding tasks due to its quality and reliability. It’s described as often more “to the point” than GPT models and excellent at understanding large codebases and following complex instructions.
  • Architectural Planning: Claude models can be suitable for this.
• Considerations: Opus can be more expensive. They may occasionally refuse to perform a task that GPT-4 would attempt.

GPT Models (GPT-3.5, GPT-4, GPT-4o, o3)

• Strengths & Ideal Use Cases:
  • Boilerplate Code: GPT-3.5 is faster for generating boilerplate.
  • Nuanced Refactoring and Logic-Heavy Components: GPT-4 is good for these.
  • Deep Architectural Restructuring and Design Patterns: GPT models are strong performers for these.
  • Simple Debugging Tasks: Effective for straightforward debugging.
  • Complex Logical Reasoning and Brainstorming: Strong all-around performers.
  • Factual Questions: Less likely to “hallucinate” on factual questions.
  • API Documentation: GPT-4o can be used for this.
  • General Default: GPT-4o balances speed and capability and is often considered a good default choice.
• Considerations: Can sometimes be “lazy” and provide incomplete code. GPT-4/GPT-4o can be more expensive than Sonnet.

Gemini Models (Gemini 2.5 Pro)

• Strengths & Ideal Use Cases:
  • Complex Design Work: Noted for its effectiveness in this area.
  • Deep Bug-Fixing: Useful for intricate bugs requiring robust reasoning capabilities.
  • Rapid Code Completions: Can be utilized for this purpose.

DeepSeek Coder

• Strengths & Ideal Use Cases:
  • Troubleshooting and Problem Analysis: Good at analyzing problems and proposing solutions.
• Considerations: Often a cost-effective and more private option as it’s hosted by Cursor. May not be as strong as frontier models for complex generation.

Local Models (e.g., Llama2 derivatives)

• Strengths & Ideal Use Cases:
  • Maximum Privacy: Your code never leaves your machine. Ideal for proprietary or highly sensitive code.
• Considerations: Generally slower and limited to smaller models compared to cloud-based options.

General Tips for Model Selection

• Balance Trade-offs: Model selection requires balancing quality, latency, and cost to suit your use case.
• Task Complexity: For easier or boilerplate tasks, consider using a lighter-weight and faster model. Save more powerful and expensive models for complex problems, such as nuanced refactoring or logic-heavy components.
• Iterative Approach: If a model gets stuck or provides unsatisfactory results, try switching to a different model. Users often swap between GPT-4o and Claude in such cases.
• “Auto” Mode: Cursor offers an “Auto” mode that intelligently selects the most appropriate premium model for a given query based on its complexity, current performance, and server reliability.
• Personal API Keys: For greater control over costs or to leverage enterprise-specific plans, you can configure your own API keys for services like OpenAI or Anthropic. However, remember that requests are still proxied through Cursor’s infrastructure.
• Cost Management: Be mindful of usage quotas, especially with premium models. Heavy use can lead to unexpectedly high bills.