During a recent GPT-5 livestream, OpenAI presented charts showcasing the model's capabilities. However, inconsistencies in the scales of some graphs have raised concerns. One chart, intended to demonstrate GPT-5's performance in 'deception evals,' displayed a misleading scale, particularly in 'coding deception' results. While the onstage chart indicated a 50.0% deception rate for GPT-5, a smaller model, o3, was shown with a larger bar despite a lower score of 47.4%. OpenAI's GPT-5 blog post presents different figures, labeling GPT-5's deception rate as 16.5%, suggesting potential errors in the livestream presentation.