During the great live broadcast, the GPT-5 on Thursday Opeli showed several charts that made the model seem impressive-but if you look carefully, some charts were somewhat turned off.
In one ironically, showing how well the GPT-5 is done in “fraud evolving in different models”, the scale is everywhere. For example, in the case of “coding fraud”, the chart shown on the stage says that GPT-5 apparently receives 50.0 percent of fraud with thinking, but this is compared to the lower result of O3 O3 Openai, which somehow has a larger belt. Opeli seems to have exact numbers for this chart Post on the GPT-5 blogHowever, where the GPT-5 fraud indicator is marked as 16.5 percent.
WITH this chartOpeli showed on stage that one of the GPT-5 results is lower than O3, but it is shown with a larger belt. On the same chart, O3 and GPT-4O results are different, but shown with equally size bars. It was bad enough that the CEO of Altman himself commented on calling it “Mega Chart Screpup”, although he noticed that the correct version is in the OpenAI blog.
Also an Openai marketing worker He apologizedSaying: “We repaired the blog chart, we apologize for the unintentional crime with charts.”
On Friday, in response to the Reddit user, asking about charts, Altman He said it “The numbers here were accurate, but we fucked up the bollard charts in live broadcast overnight; on the next slide we fucked up the numbers.” He also noticed that the blog post and the system card were “accurate” and said that “people worked late and were very tired, and the human error bothers. In the last hours a lot is live.”
This is still not a great appearance of the company on a great day of premiere – especially when it advertises “significant progress in reducing hallucinations” using a recent model.
Update, August 8: Reddit comment from Altman has been added.