OpenAI's livestream for GPT-5 displayed several inaccurate charts, leading to criticism. A chart meant to show coding deception rates inaccurately represented GPT-5's rate of 50% as being smaller than the rate for model o3, which is lower at 47.4%. In another example, a comparison chart portrayed two different percentages (69.1% and 30.8%) with identical bar sizes, while a third (52.8%) incorrectly appeared larger. Following the errors, OpenAI's CEO publicly acknowledged the mistakes, calling it a 'mega chart screwup' and apologizing for the 'unintentional chart crime.'
Several charts included in OpenAI's GPT-5 livestream on Thursday had some clear mistakes. One chart compared the models GPT-5 with thinking and OpenAI o3 on a metric called "coding deception." The former model has a deception rate of 50%, but this bar was less than half the size of the bar for o3, which has a smaller deception rate of 47.4%.
Another chart comparing models on a different metric depicted 69.1% and 30.8% with the same size bar. A bar for 52.8% was also larger than both of these, though that's smaller than 69.1%.
OpenAI CEO Sam Altman tweeted, "wow a mega chart screwup from us earlier." A marketing employee for OpenAI tweeted that the issue had been fixed in the blog post of the announcement, apologizing for the "unintentional chart crime."
Collection
[
|
...
]