The language model competition in the AI ​​field is becoming increasingly fierce, especially in programming assistance.

Gemini 2.5 Pro and Claude 3.7 Sonnet are leaders in this field. This article compares and analyzes the coding capabilities of the two through a series of programming tests and benchmark evaluations.

Core conclusions:

• Gemini 2.5 Pro slightly beat Claude 3.7 Sonnet's 62.3% in the SWE Bench hardcore programming test with a pass rate of 63.8%.

• Both models have their own strengths and limitations when completing different types of programming tasks. Gemini 2.5 Pro is often more accurate and faster when generating code solutions.

• The actual choice depends on the project requirements and the specific type of programming task.

Gemini 2.5 Pro Overview

Despite its release date, Gemini 2.5 Pro continues to generate buzz with its upgraded reasoning analysis capabilities, which were previously available only to Gemini Advanced subscribers but are now available to the public for free.

Although it is a newcomer, Gemini 2.5 Pro has surpassed rivals such as ChatGPT 4 in some tests (except for programming and multi-round conversations).

Surprisingly, in the "Ultimate Human Exam" test, its version without web search function enabled surpassed OpenAI's deep research model, achieving amazing results.

Coding Challenge Test

To evaluate the programming capabilities of Gemini 2.5 Pro and Claude 3.7 Sonnet, we tested both models with a series of coding tasks. The results are summarized below:

1. Flight Simulator

Requirements: Use JavaScript to develop a simple flight simulator that includes a basic aircraft model that can take off from a flat runway. The aircraft movement must be controlled by keyboard input (such as arrow keys or WASD keys), and a basic city landscape must be generated with block buildings similar to the style of "Minecraft".

Gemini 2.5 Pro Performance:

Successfully generated a working flight simulator code. The generated code is completely correct, the aircraft controls are smooth, and the city landscape is rendered accurately.

const plane = document.createElement( 'div' );
plane.style.position =  'absolute' ;
plane.style.left =  '50%' ;
plane.style.bottom =  '10px' ;
plane.style.width =  '50px' ;
plane.style.height =  '20px' ;
plane.style.background =  'gray' ;
document.body.appendChild(plane);


document.addEventListener( 'keydown' , (event) => {
    if  (event.key ===  'ArrowUp' ) {
        plane.style.bottom = `${parseInt(plane.style.bottom) +  10 }px`;
    }
    if  (event.key ===  'ArrowDown' ) {
        plane.style.bottom = `${parseInt(plane.style.bottom) -  10 }px`;
    }
});

Claude 3.7 Sonnet Performance:

The generated sample code was of poor quality, with issues ranging from poor aircraft handling to poor city rendering. Even following the prompts to optimize did not improve anything.

const plane = document.createElement( 'div' );
plane.style.position =  'fixed' ;
plane.style.left =  '45%' ;
plane.style.bottom =  '5%' ;
plane.style.width =  '60px' ;
plane.style.height =  '25px' ;
plane.style.backgroundColor =  'blue' ;
document.body.appendChild(plane);


document.addEventListener( 'keydown' , (event) => {
    switch(event.key) {
        case  'w' : plane.style.bottom = `${parseInt(plane.style.bottom) +  15 }px`;  break ;
        case  's' : plane.style.bottom = `${parseInt(plane.style.bottom) -  15 }px`;  break ;
    }
});

Test conclusion: In this test, Gemini 2.5 Pro beats Claude 3.7 Sonnet with a more realistic and accurate solution.

2. Rubik's Cube Solver

Topic: Write JavaScript code to display and solve a 3x3 Rubik's Cube. The code should display the Rubik's Cube in 3D and show the solution step by step.

Gemini 2.5 Pro's solution:

The complete code was successfully generated, the 3D Rubik's Cube model was accurately presented, and the correct solution steps were demonstrated.

from  rubik_solver  import  utils
cube_state =  'RRRRRRRRRBBBBBBBBBOOOOOOOGGGGGGGGGYYYYYYYYYWWWWWWWWW'
solution = utils.solve(cube_state,  'Kociemba' )
print(solution)

Claude 3.7 Sonnet performance:

The solution given was incorrect, not only was the color incorrect, but the cube was also not solved correctly. Even after multiple additional hints, the model was unable to correct these errors.

def solve_rubiks_cube () : 
    moves = [ 'U' ,  'R' ,  "L" ,  "B" ,  "F" ,  "D" ]
    solution = []
    for  i  in  range( 10 ):
        solution.append(random.choice(moves))
    return  solution


print(solve_rubiks_cube())

Test conclusion: Gemini 2.5 Pro performed better in this round of testing, providing correct and visually excellent graphical solutions.

Performance Benchmarks

Benchmark test data shows that the Gemini 2.5 Pro leads in all but one category.

It surpassed Claude 3.7 Sonnet by 30% in the AIME test and won the GPQA test with a score of 84% to 68%.

A Fortune 500 transportation company deployed Gemini 2.5 Pro for route optimization in March 2025, and the report showed a 15% reduction in fuel consumption, a 22% increase in on-time delivery rate, and an average annual cost savings of US$3.5 million.

Final Conclusion

In the field of programming assistants, both Gemini 2.5 Pro and Claude 3.7 Sonnet have demonstrated outstanding capabilities.

Both are top-notch, but Gemini 2.5 Pro has an advantage when dealing with difficult programming problems and exam scenarios. Its excellent performance makes it an ideal choice for developers who need top-notch programming assistance.

This article is synchronized from Knowledge Planet "AI Disruption"

I am a top editor at Substack and Medium. I also work as an independent developer.

The planet shares AI trends and foreign digital marketing.

Planet is not free. The price is 99 yuan/year, 0.27 yuan/day.

• First, it has operating costs, and I hope it can be self-closed so that it can operate stably in the long term;

• The second is the selection of people. I don’t want a mixed bag of people. I hope to find people who are interested in and passionate about AI.