Скачать или смотреть Gemini 2.5 pro Project Mariner Finished a 30-Min Task in 4

Gemini 2.5 Pro 05-06 has generated considerable attention following its release, especially among users building complex coding projects. This detailed evaluation compares the model against its earlier version and major competitors like GPT-4.1 and Claude 3.7 Sonnet, focusing on a single prompt designed to simulate a physics-driven dice-in-a-box scenario.

The test prompt asked models to create a Python program where dice fall into a container under gravity, collide realistically, and respond to the box tilting back and forth. Each model was assessed on criteria such as physics realism, box tilt behavior, dice stacking, and output stability.

Gemini 2.5 Pro 05-06 improved over its preview version in physics fidelity and visual appeal but failed to render a proper box or dice. GPT-4.1 included the tilting box but delivered a chaotic simulation with overly aggressive motion. Claude 3.7 Sonnet produced accurate dice visuals in its standard mode, though it lacked proper movement and a defined box. The extended thinking mode introduced creative elements and interaction but drifted away from the prompt's requirements.

Despite advancements, none of the models fully satisfied all criteria. Gemini 2.5 Pro 05-06 and Claude 3.7 Sonnet both performed well in isolated aspects but fell short of delivering a complete, prompt-accurate simulation. GPT-4.1, while not a reasoning model, managed a reasonable but turbulent execution.

The breakdown offers insights into how different large language models interpret and execute the same coding task, highlighting the strengths and limitations of each under practical, prompt-specific conditions.