Gemini 2.5 Pro Extended Thinking Mode

Setup and context

Gemini 2.5 Pro's Extended Thinking is a feature that deepens the AI's reasoning process, generating more accurate and well-founded answers to complex problems. Rather than providing instant responses, it takes time to analyze problems step-by-step, considers multiple approaches, and derives the final answer through comprehensive evaluation.

What is Extended Thinking?

Extended Thinking mode allows Gemini 2.5 Pro to perform an "invisible reasoning process" for a given problem. This process follows these steps:

1. Problem Analysis: Break down the given problem into components
2. Hypothesis Generation: Consider multiple approaches
3. Reasoning: Verify the pros and cons of each approach
4. Conclusion: Generate the optimal answer

This entire process is exposed to API users as "thinking" and can be referenced alongside the final answer.

Differences from Standard Mode

Applicable Use Cases

Extended Thinking is particularly effective for complex problem-solving scenarios:

✓ Ideal Use Cases

• Mathematical & Physics Calculations: Multi-step computations and proofs
• Code Generation & Debugging: Complex algorithm implementation
• Logic Problems: Puzzles and constraint satisfaction problems
• Technical Design: Architecture decisions and best practice selection
• Data Analysis: Statistical analysis and trend prediction

✗ Not Suitable For

• Template or boilerplate generation
• Simple information retrieval
• Real-time chat responses

Basic API Usage

Step 1: Environment Setup

pip install google-genai

Step 2: Basic Extended Thinking Request

from google import genai
 
# Initialize client
client = genai.Client(api_key="YOUR_API_KEY")
 
# Solve a problem using Extended Thinking mode
response = client.models.generate_content(
    model="gemini-2.5-pro",
    contents="Solve this equation: 2x² + 5x - 3 = 0",
    config=genai.types.GenerateContentConfig(
        thinking={
            "type": "enabled",
            "budget_tokens": 5000,  # Max tokens for thinking
        }
    ),
)
 
# Display thinking process and final answer
for part in response.content.parts:
    if part.thinking:
        print("【Thinking Process】")
        print(part.thinking)
    elif part.text:
        print("【Final Answer】")
        print(part.text)

Step 3: Solving More Complex Problems

# Generate efficient code for computing Fibonacci numbers
response = client.models.generate_content(
    model="gemini-2.5-pro",
    contents="""
Write Python code meeting these requirements:
1. Function to compute the nth Fibonacci number
2. Achieve O(log n) time complexity
3. Use matrix exponentiation
4. Include test cases
    """,
    config=genai.types.GenerateContentConfig(
        thinking={
            "type": "enabled",
            "budget_tokens": 8000,
        }
    ),
)
 
for part in response.content.parts:
    if part.thinking:
        print("🤔 Thinking Process:")
        print(part.thinking[:500] + "..." if len(part.thinking) > 500 else part.thinking)
    elif part.text:
        print("\n💻 Generated Code:")
        print(part.text)

Practical Example: Complex Algorithm Problem

Solve the classic "3Sum" problem using Extended Thinking:

response = client.models.generate_content(
    model="gemini-2.5-pro",
    contents="""
Write efficient Python code to solve the 3Sum problem:
- Input: Integer array and target sum
- Output: All unique triplets that sum to target
- Time complexity: O(n²) or better
- Include test cases
 
Example: arr = [-1, 0, 1, 2, -1, -4], target = 0
Expected: [[-1, -1, 2], [-1, 0, 1]]
    """,
    config=genai.types.GenerateContentConfig(
        thinking={
            "type": "enabled",
            "budget_tokens": 10000,
        }
    ),
)
 
print("=== 3Sum Problem Solution ===\n")
for part in response.content.parts:
    if part.thinking:
        print("Thinking Process (Overview):")
        lines = part.thinking.split('\n')
        print('\n'.join(lines[:10]))
        print(f"... ({len(lines)} lines total)")
    elif part.text:
        print("\nFinal Code:\n")
        print(part.text)

Parameter Details

thinking Object

Larger budget_tokens values enable deeper thinking but increase response time and cost. Recommended range: 5,000–15,000.

Best Practices

1. Enable only for complex problems: Use standard mode for simple questions
2. Set budget_tokens appropriately: Adjust within 5,000–15,000 based on problem complexity
3. Leverage thinking output: Review the reasoning to understand answer justification
4. Implement error handling: Account for potential timeouts with retry logic

Conclusion

Gemini 2.5 Pro's Extended Thinking mode significantly improves AI accuracy and reliability for complex problem-solving. Whether you're tackling mathematical calculations, code generation, or logic problems, Extended Thinking delivers well-founded, high-precision answers. The API is straightforward—just configure the thinking parameter.

Try Extended Thinking in your projects today!