Code Execution

Code execution enables Gemini to generate, run, and iteratively refine Python code to solve problems. The model can write code, execute it in a secure sandbox, observe results, and automatically fix errors until producing a working solution.

Why Code Execution?

Iterative Learning

Model observes execution results and refines code automatically

Mathematical Accuracy

Solve complex calculations with verified results

Data Processing

Generate and test data analysis code

Supported Models

Code execution is available on:

gemini-3.1-pro-preview
gemini-3-flash-preview
gemini-2.5-pro
gemini-2.5-flash

Basic Example

Setup

from google import genai
from google.genai.types import (
    Tool,
    ToolCodeExecution,
    GenerateContentConfig
)

client = genai.Client(
    vertexai=True,
    project=PROJECT_ID,
    location=LOCATION
)

# Define code execution tool
code_execution_tool = Tool(code_execution=ToolCodeExecution())

Generate and Execute Code

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="What is the sum of the first 50 prime numbers? Generate and run code.",
    config=GenerateContentConfig(
        tools=[code_execution_tool],
        temperature=0
    )
)

print(response.text)
# Output: The sum of the first 50 prime numbers is 5117.

View Generated Code

Access the Python code that was generated:

for part in response.candidates[0].content.parts:
    if part.executable_code:
        print("Generated Code:")
        print(part.executable_code.code)

Output:

def is_prime(n):
    if n <= 1:
        return False
    if n <= 3:
        return True
    if n % 2 == 0 or n % 3 == 0:
        return False
    i = 5
    while i * i <= n:
        if n % i == 0 or n % (i + 2) == 0:
            return False
        i += 6
    return True

primes = []
num = 2
while len(primes) < 50:
    if is_prime(num):
        primes.append(num)
    num += 1

sum_of_primes = sum(primes)
print(f'{sum_of_primes=}')

View Execution Results

Access code execution output and status:

for part in response.candidates[0].content.parts:
    if part.code_execution_result:
        print(f"Output: {part.code_execution_result.output}")
        print(f"Outcome: {part.code_execution_result.outcome}")

Output:

Output: sum_of_primes=5117
Outcome: Outcome.OUTCOME_OK

Mathematical Problem Solving

Complex Calculations

prompt = """
A bat and a ball cost $1.10 in total. 
The bat costs $1.00 more than the ball.
How much does the ball cost?
Solve this step-by-step with code.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool],
        temperature=0
    )
)

print(response.text)

Data Analysis

prompt = """
Generate random sales data for 12 months and:
1. Calculate total annual revenue
2. Find the best performing month
3. Calculate month-over-month growth rate
4. Identify any trends

Use code to analyze the data.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)

Code Execution in Chat

Maintain context across multiple code-related queries:

chat = client.chats.create(
    model="gemini-3-flash-preview",
    config=GenerateContentConfig(
        tools=[code_execution_tool],
        temperature=0
    )
)

# First query
response = chat.send_message(
    "Write a function that checks if a year is a leap year."
)
print("Assistant:", response.text)

# Follow-up with context
response = chat.send_message(
    "Now write unit tests for that function."
)
print("\nAssistant:", response.text)

# Another follow-up
response = chat.send_message(
    "Test it with the years 2000, 2020, 2024, and 1900."
)
print("\nAssistant:", response.text)

Streaming Code Execution

Stream responses as code is generated and executed:

for chunk in client.models.generate_content_stream(
    model="gemini-3-flash-preview",
    contents="Calculate the factorial of 20. Show your work with code.",
    config=GenerateContentConfig(
        tools=[code_execution_tool],
        temperature=0
    )
):
    print(chunk.text, end="")

Data Exploration

Analyze datasets iteratively:

prompt = """
I have the following sales data:
- January: $45,000
- February: $52,000
- March: $48,000
- April: $61,000
- May: $58,000
- June: $63,000

Analyze this data:
1. Calculate average monthly sales
2. Find percentage growth from January to June
3. Identify which month had the highest growth rate
4. Project sales for July based on the trend

Use Python code to perform the analysis.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)

Algorithmic Problem Solving

Sorting Algorithm

prompt = """
Implement the quicksort algorithm in Python.
Test it with the array [64, 34, 25, 12, 22, 11, 90]
Verify the result is correctly sorted.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)

Graph Algorithms

prompt = """
Implement Dijkstra's shortest path algorithm.
Create a sample graph with 5 nodes and find the shortest
path from node A to node E.
Show the path and its total distance.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)

Statistical Analysis

prompt = """
Generate 1000 random data points from a normal distribution
with mean=100 and std=15.

Then:
1. Calculate mean, median, and standard deviation
2. Find the 95th percentile
3. Count how many values are within 1 standard deviation
4. Verify this matches the 68-95-99.7 rule

Use Python code with numpy.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)

Code Validation

The model automatically validates and fixes code:

prompt = """
Write a function to calculate fibonacci numbers recursively.
Test it with n=10 and n=20.
If there are any issues, fix them.
"""

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=prompt,
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)
# Model will generate code, test it, and refine if needed

Extract Code and Results

Programmatically access generated code and results:

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="Calculate pi using the Monte Carlo method with 10000 samples.",
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

code = None
result = None

for part in response.candidates[0].content.parts:
    if part.executable_code:
        code = part.executable_code.code
    if part.code_execution_result:
        result = part.code_execution_result.output
        outcome = part.code_execution_result.outcome

print("Generated Code:")
print(code)
print("\nExecution Result:")
print(result)
print(f"\nOutcome: {outcome}")

Combine with Other Features

Code Execution + Function Calling

from google.genai.types import FunctionDeclaration

get_data = FunctionDeclaration(
    name="get_sales_data",
    description="Retrieve sales data from database",
    parameters={"type": "object", "properties": {}}
)

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="Get the sales data and analyze trends using Python.",
    config=GenerateContentConfig(
        tools=[
            code_execution_tool,
            Tool(function_declarations=[get_data])
        ]
    )
)

Code Execution + Multimodal

from google.genai.types import Part

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents=[
        Part.from_uri(
            file_uri="gs://samples/data-chart.png",
            mime_type="image/png"
        ),
        "Extract the data points from this chart and calculate the trend line."
    ],
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

print(response.text)

Best Practices

Clear Instructions

Specify exactly what the code should accomplish

Temperature 0

Use temperature=0 for deterministic code generation

Test Cases

Include test cases in your prompts

Verify Results

Always validate code execution outcomes

Effective Prompts

✅ Good Prompts:

"Calculate the sum of prime numbers less than 100. Generate and run code."
"Write a function to reverse a string. Test it with 'hello world'."
"Analyze this dataset and find outliers using statistical methods."

❌ Less Effective:

"Do some math"  # Too vague
"Write code"    # No clear objective
"Calculate"     # Missing specifics

Limitations

Code Execution Limitations:

Python standard library only (no pip packages)
Limited execution time (typically 30 seconds)
No file system access
No network access
Memory limits apply

Available Libraries

The execution environment includes Python standard library modules:

math, random, statistics
datetime, time
json, csv
itertools, functools
collections, heapq
And other standard library modules

Error Handling

response = client.models.generate_content(
    model="gemini-3-flash-preview",
    contents="Your code request",
    config=GenerateContentConfig(
        tools=[code_execution_tool]
    )
)

for part in response.candidates[0].content.parts:
    if part.code_execution_result:
        outcome = part.code_execution_result.outcome
        
        if outcome == "OUTCOME_OK":
            print("Code executed successfully")
            print(f"Result: {part.code_execution_result.output}")
        elif outcome == "OUTCOME_FAILED":
            print("Code execution failed")
            print(f"Error: {part.code_execution_result.output}")
        elif outcome == "OUTCOME_DEADLINE_EXCEEDED":
            print("Code execution timed out")

Debugging Code

Ask the model to debug and fix issues:

prompt = """
This code has a bug:

```python
def fibonacci(n):
    if n <= 0:
        return 0
    if n == 1:
        return 1
    return fibonacci(n-1) + fibonacci(n-2)

print(fibonacci(50))  # This takes forever!

Find the problem and provide an optimized solution. Test both versions. """ response = client.models.generate_content( model=“gemini-3-flash-preview”, contents=prompt, config=GenerateContentConfig( tools=[code_execution_tool] ) ) print(response.text)

## Next Steps

<CardGroup cols={2}>
  <Card title="Function Calling" icon="function" href="/gemini/function-calling">
    Combine code execution with external APIs
  </Card>
  
  <Card title="Multimodal" icon="images" href="/gemini/multimodal">
    Generate code from images and diagrams
  </Card>
  
  <Card title="Context Caching" icon="database" href="/gemini/context-caching">
    Cache large code contexts
  </Card>
  
  <Card title="Grounding" icon="anchor" href="/gemini/grounding">
    Ground code generation in documentation
  </Card>
</CardGroup>

## Resources

- [Code Execution Documentation](https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/code-execution)
- [Python Standard Library](https://docs.python.org/3/library/)
- [Best Practices Guide](https://cloud.google.com/vertex-ai/generative-ai/docs/learn/prompts/code-generation-prompts)

Getting Started

Gemini Models

Agents

RAG & Search

Embeddings & Vector Search

Vision

Audio

​Code Execution

​Why Code Execution?

Iterative Learning

Mathematical Accuracy

Data Processing

​Supported Models

​Basic Example

​Setup

​Generate and Execute Code

​View Generated Code

​View Execution Results

​Mathematical Problem Solving

​Complex Calculations

​Data Analysis

​Code Execution in Chat

​Streaming Code Execution

​Data Exploration

​Algorithmic Problem Solving

​Sorting Algorithm

​Graph Algorithms

​Statistical Analysis

​Code Validation

​Extract Code and Results

​Combine with Other Features

​Code Execution + Function Calling

​Code Execution + Multimodal

​Best Practices

Clear Instructions

Temperature 0

Test Cases

Verify Results

​Effective Prompts

​Limitations

​Available Libraries

​Error Handling

​Debugging Code

Build docs developers (and LLMs) love

Code Execution

Why Code Execution?

Supported Models

Basic Example

Setup

Generate and Execute Code

View Generated Code

View Execution Results

Mathematical Problem Solving

Complex Calculations

Data Analysis

Code Execution in Chat

Streaming Code Execution

Data Exploration

Algorithmic Problem Solving

Sorting Algorithm

Graph Algorithms

Statistical Analysis

Code Validation

Extract Code and Results

Combine with Other Features

Code Execution + Function Calling

Code Execution + Multimodal

Best Practices

Effective Prompts

Limitations

Available Libraries

Error Handling

Debugging Code