I have a hard time believing this isn't made up given OpenAI Codex performance on my tasks.