OpenAI’s live demo of GPT-5 last night included graphs with errors that users immediately pointed out. A comparison chart showing how accurate GPT-5 was compared to OpenAI’s older AI models, o3 and GPT-4o, displayed 52.8% accuracy (with thinking) as higher than o3’s 69.1%. 

Meanwhile, o3’s 69.1% accuracy was the same level as GPT-4o’s 30.8% on the bar graph. 

Another bar graph showing “Deception Evals across models” showed GPT-5 with thinking attaining 50% but still having a much smaller bar than o3’s 47.4%. 

After several users on X pointed out the errors, CEO Sam Altman responded to them saying, “wow a mega chart screwup from us earlier.” He noted that the same chart was accurate on the blog post for the release. 

Users speculated that OpenAI had used their own AI models to generate these graphs but OpenAI hasn’t clarified if that was the case. 

Altman has said called GPT-5 a “PhD-level expert” unlike the previous flagship models which felt more like speaking with a “student.”