Model Selection

Learn to select the right AI model for each task. This 10-minute guide will help you balance speed, cost, and capability to maximize productivity while controlling expenses.

Model Landscape Overview

Tip

Key Principle: Use the smallest model that can reliably complete your task. Reserve premium models for situations where their unique capabilities justify the cost.

Quick Reference Matrix

Model	Speed	Cost	Context	Best Use Case
Claude Sonnet 4.5	⚡⚡⚡	$	1M	Daily coding (95%+ of tasks) - best model
gpt-5-codex	⚡⚡⚡	$$$	200k	Bug fixing & UI generation (Cursor)
Claude Opus 4	⚡⚡	$$$$$	200k	Ultra-complex architecture (rare)
Gemini 2.5 Pro	⚡⚡⚡	$$	1M	Extreme context needs (>1M tokens)

Cost Comparison

Monthly Budget Impact

Typical Developer Usage

• Sonnet 4.5 only: ~$50/month
• Sonnet 4.5 + gpt-5-codex: ~$100/month
• With Opus 4: ~$200/month (rarely needed)
• Optimized: ~$50-100/month

Token Economics

Per Million Tokens

• Sonnet 4.5: $3 (input) / $15 (output)
• Opus 4: $15 / $75 (5x Sonnet)
• gpt-5-codex: Premium (Cursor)
• Gemini 2.5 Pro: $2 / $10

Claude Sonnet 4.5 - The Best Daily Driver

Released: September 29, 2025 (announcement)

When to Use Sonnet 4.5 (Default)

Perfect For

• All daily coding tasks (90%+ of work)
• Feature implementation
• Bug fixes (though gpt-5-codex specializes in this)
• Code reviews
• Large-scale refactoring (1M context!)
• Test writing
• Documentation
• API integration
• Database queries
• Architecture design
• Agent building

Why It's Better

• 1 million token context - holds entire codebases
• Best coding model in the world
• Superior to Opus 4 for most tasks
• State-of-the-art on SWE-bench
• Can focus for 30+ hours on complex tasks
• Same cost as previous Sonnet ($3/$15)
• Best at building complex agents

Sonnet 4.5 Optimization Tips

1. Clear, specific prompts yield better results than vague requests
2. Include examples in your prompts for consistent output
3. Break complex tasks into smaller, focused requests
4. Use project rules to maintain consistency

Real-World Examples

// EXCELLENT: Take advantage of 1M context with Sonnet 4.5
"Analyze the entire authentication system across all files,
identify security vulnerabilities, and refactor to use
industry best practices. Include tests for all changes."

// GOOD: Specific request for Sonnet 4.5
"Create a TypeScript function that validates email addresses
using regex, returns a Result<string, ValidationError> type,
and includes unit tests"

gpt-5-codex - Bug Fixing & UI Specialist

Available in: Cursor only

When to Use gpt-5-codex

Decision Framework

Use gpt-5-codex when you need:

1. Bug Fixing: Specialized training for identifying and fixing bugs quickly
2. UI Generation: Excellent at creating and refining user interfaces
3. Frontend Work: Strong understanding of React, Vue, Angular, and modern CSS
4. Quick Iterations: Fast at UI tweaks and visual refinements

Tips for gpt-5-codex

1. Use for targeted bug fixes - especially production issues
2. Great for component creation - UI components, forms, layouts
3. Combine with Sonnet 4.5 - use codex for bugs/UI, Sonnet for architecture
4. Fast iterations - quick feedback loop for visual work

Example Workflow

// Use gpt-5-codex for bug fixing
"This button isn't responding to clicks on mobile.
Debug and fix the touch event handlers."

// Use gpt-5-codex for UI generation
"Create a responsive navigation menu with dropdown support,
mobile hamburger menu, and smooth animations."

// Switch to Sonnet 4.5 for broader work
"Now integrate this navigation into our routing system
and add proper authentication checks."

Claude Opus 4 - Rarely Needed Now

Caution

Reality Check: With Sonnet 4.5’s capabilities, you’ll rarely need Opus 4. It’s 5x the cost and Sonnet 4.5 handles most “complex” tasks better.

When to Upgrade to Opus 4 (Rare Cases)

Decision Framework

Use Opus 4 only when:

1. Ultra-Complex Architecture: System design that truly stumps Sonnet 4.5 (rare)
2. Specific Deep Reasoning: When Sonnet 4.5 repeatedly fails (very rare)

Note: Try Sonnet 4.5 first. 95% of the time it’s enough, even for complex tasks.

Opus 4 Usage Patterns

1. Start with Planning

   "Analyze our current authentication system and propose
   a migration plan to OAuth 2.0 with backward compatibility"

2. Generate Architecture

   "Design a scalable event-driven architecture for our
   notification system supporting email, SMS, and push"

3. Complex Problem Solving

   "Optimize this graph traversal algorithm for finding
   shortest paths in a weighted directed graph with
   negative edges"

Gemini 2.5 Pro - For Extreme Context Only

Caution

When to Use: Only when your task genuinely exceeds Claude Sonnet 4.5’s 1M token context. This is extremely rare.

Rare Use Cases

Use Gemini 2.5 Pro only when:

• Your codebase genuinely exceeds 1M tokens (very large monorepos)
• You need to analyze absolutely everything at once
• Sonnet 4.5’s 1M context isn’t enough (rare)

Reality: Sonnet 4.5’s 1M context handles most entire codebases. You probably won’t need Gemini 2.5 Pro.

Model Selection Decision Tree

graph TD Start[New Task] --> Type{Task Type?} Type -->|Bug Fix or UI| Codex[gpt-5-codex in Cursor] Type -->|Everything Else| Sonnet[Claude Sonnet 4.5] Codex --> Working{Working well?} Sonnet --> Working2{Working well?} Working -->|Yes| Done[Continue] Working -->|No| Sonnet Working2 -->|Yes| Done Working2 -->|No| Check{Exceeds 1M tokens?} Check -->|Yes| Gemini[Try Gemini 2.5 Pro] Check -->|No| Opus[Try Opus 4 - very rare]

Cost Optimization Strategies

1. Progressive Enhancement

1. Start with Sonnet 4.5 for 90%+ of all work (it’s the best!)
2. Use gpt-5-codex for bug fixes and UI generation
3. Use o3 only for extremely complex debugging
4. Use Gemini only if you exceed 1M token context

2. Context Management

Reduce Context

• Clear chat regularly
• Use focused @mentions
• Exclude irrelevant files
• Summarize long discussions

Reuse Context

• Save useful prompts
• Create project rules
• Build prompt templates
• Use memory feature

3. Batch Operations

# EXPENSIVE: Multiple Opus 4 calls
"Refactor the auth service" (Opus 4)
"Now refactor the user service" (Opus 4)
"Now refactor the profile service" (Opus 4)

# EFFICIENT: Single comprehensive call
"Refactor auth, user, and profile services to follow
our new architecture pattern. Provide implementation
plan first, then execute." (Opus 4 once)

Model Switching Workflow

Quick Switch Method

Tip

Use keyboard shortcut Ctrl + . to quickly switch models mid-conversation

Strategic Switching Pattern

1. Exploration Phase (Sonnet 4.5)

   • Understand the problem
   • Gather context
   • Initial attempts

2. Planning Phase (Opus 4)

   • Architecture design
   • Break down complex tasks
   • Create implementation plan

3. Implementation Phase (Sonnet 4.5)

   • Execute the plan
   • Write code
   • Create tests

4. Debug Phase (o3 if needed)

   • Solve complex issues
   • Optimize algorithms
   • Fix edge cases

Real-World Scenarios

Scenario 1: New Feature

Task: Add user notifications
Models Used:
  - Sonnet 4.5: Initial implementation (90%)
  - Opus 4: System design (10%)
Total Cost: ~$5
Time Saved: 4 hours

Scenario 2: Performance Issue

Task: Fix memory leak in production
Models Used:
  - Sonnet 4.5: Initial investigation (20%)
  - o3: Deep analysis and fix (80%)
Total Cost: ~$15
Time Saved: 8 hours debugging

Scenario 3: Large Refactor

Task: Migrate to new framework
Models Used:
  - Opus 4: Planning (20%)
  - Gemini 2.5: Analysis (30%)
  - Sonnet 4.5: Implementation (50%)
Total Cost: ~$40
Time Saved: 20 hours

Advanced Model Techniques

1. Model Chaining

Opus 4: "Create a detailed plan for implementing OAuth"
  ↓ (Save plan to file)
Sonnet 4.5: "Implement step 1 from oauth-plan.md"
Sonnet 4.5: "Implement step 2 from oauth-plan.md"

2. Parallel Model Usage

Open multiple Cursor instances:

• Instance 1: Opus 4 for architecture
• Instance 2: Sonnet 4.5 for implementation
• Instance 3: o3 for testing edge cases

3. Model-Specific Prompting

Sonnet 4.5 Prompts

- Direct and specific
- Include code examples
- Reference file paths
- Clear success criteria

Opus 4 Prompts

- High-level goals
- Ask for reasoning
- Request alternatives
- Include constraints

o3 Prompts

- Provide all context
- Include error logs
- Ask for step-by-step
- Request verification

Monitoring Model Usage

Track Your Costs

1. Check usage: Settings → Usage
2. Set budget alerts
3. Review weekly patterns
4. Optimize based on data

Usage Patterns to Watch

• High Opus 4 usage: Consider better planning
• Repeated similar tasks: Create rules/templates
• Long conversations: Clear context more often
• Failed attempts: Switch models earlier

Model Selection Checklist

Before starting a task, ask:

• Can Sonnet 4.5 handle this? (Yes, 90%+ of the time!)
• Is this a bug fix or UI work? (Consider gpt-5-codex)
• Do I need > 1M tokens? (Consider Gemini 2.5 Pro)
• Am I using the right prompting style?
• Is the model working well or should I switch?

Next Steps

Continue to Project Rules

Now let’s set up project rules to ensure consistent AI behavior across all models.

Project Rules →

Time: 10 minutes

Tip

Key Takeaway: Claude Sonnet 4.5 is the best coding model with 1M context - use it for 90%+ of tasks. Complement with gpt-5-codex for bug fixes and UI work. Check Cursor changelog for updates.