OpenAI models outperform Gemini on coding benchmarks with fewer tokens

OpenAI models achieve higher coding benchmark scores with a fraction of the tokens used by Gemini, which averages 250k per task. The video attributes this to the 'Grug speak' theory from leaked GPT 5.5 reasoning traces.