\n<\/p>\n
During its big GPT-5 livestream on Thursday<\/a>, OpenAI showed off a few charts that made the model seem quite impressive \u2014 but if you look closely, some graphs were a little bit off.<\/p>\n<\/div>\n In one, ironically showing how well GPT-5 does in \u201cdeception evals across models,\u201d the scale is all over the place. For \u201ccoding deception,\u201d for example, the chart shown onstage says GPT-5 with thinking apparently gets a 50.0 percent deception rate, but that\u2019s compared to OpenAI\u2019s smaller 47.4 percent o3 score which somehow has a larger bar. OpenAI appears to have accurate numbers for this chart in its GPT-5 blog post<\/a>, however, where GPT-5\u2019s deception rate is labeled as 16.5 percent.<\/p>\n<\/div>\n With this chart<\/a>, OpenAI showed onstage that one of GPT-5\u2019s scores is lower<\/em> than o3\u2019s but is shown with a bigger bar. In this same chart, o3 and GPT-4o\u2019s scores are different but shown with equally-sized bars. It was bad enough that CEO Sam Altman commented on it, calling it<\/a> a \u201cmega chart screwup,\u201d though he noted that a correct version is in OpenAI\u2019s blog post<\/a>.<\/p>\n<\/div>\n An OpenAI marketing staffer also apologized<\/a>, saying, \u201cWe fixed the chart in the blog guys, apologies for the unintentional chart crime.\u201d<\/p>\n<\/div>\n OpenAI didn\u2019t immediately respond to a request for comment. And while it\u2019s unclear if OpenAI used GPT-5<\/a> to actually make the charts, it\u2019s still not a great look for the company on its big launch day \u2014 especially when it is touting the \u201csignificant advances in reducing hallucinations\u201d with its new model.<\/p>\n<\/div>\n