EFFICIENCY OF CODE GENERATION BY ARTIFICIAL INTELLIGENCE GPT-4O FOR UKRAINIAN-LANGUAGE TASKS OF VARIOUS COMPLEXITY ON THE EOLYMP PLATFORM

In this study, 45 programming tasks from the Eolymp automated system (eolymp.com) were selected and categorized into three groups (A, B, C) based on difficulty (easy, medium, hard). The problem statements, constraints, and input/output examples were provided to the GPT-4o model in Ukrainian. Using a predefined prompt, the model generated C++ solutions, which were tested on Eolymp. Results showed that GPT-4o achieved the highest success rate for easy problems (54.8±44.6%), significantly outperforming medium (16.73±23.8%) and hard problems (21.4±24.7%) (p<0.017). However, no significant difference was found between medium and hard problem success rates (p>0.05). The model fully solved 46.7% of easy problems but failed all test cases for several tasks across all groups. These findings suggest that while GPT-4o performs well on simple tasks, it struggles with more complex programming problems.