Home Editor's Pick OpenAI’s new model is better at reasoning and, occasionally, deceiving