It likely isn't trained on that task so performance would be worse than frontier models specifically trained for coding.