OpenAI accused of showing misleading charts during GPT-5 launch, Sam Altman calls it ‘mega screwup’

OpenAI’s new GPT-5 model was released during a live-stream on Thursday where the company showed off many comparisons in order to certify the new model’s enhanced capabilities compared to the competition and its older models. However, during the event, the company went on to show off many charts that were simply untrue and this led to an apology from an OpenAI staffer.

In one of the charts presented during the live-stream, it shows GPT-5’s supposed prowess compared to OpenAI’s o3 in different parameters. In one of the parameters called ‘coding deception’, GPT-5 gets a 50 percent deception rate compared to o3’s 47.4 score, a pretty close call statistically speaking but the graph shows almost the opposite. Instead of showing a higher bar for GPT-5, OpenAI showed off o3 with a much bigger bar.

The company went on to correct the mistake in a subsequent blogpost but the deception rate for GPT-5 there is certified as 16.5 percent.

As if that wasn’t enough, another chart from the live-stream comparing GPT-5 to o3 and GPT-4o shows the new model with a score of 74.9 compared to 69.1 and 30.8 for the other two models respectively. While o3 and GPT-4o have a vast difference in scores, the chart shows them with almost the same length bar compared to a bigger bar for GPT-5.

OpenAI CEO Sam Altman commented on the wrong charts shown during the live-stream, calling it a “mega chart screwup” while stating that a correct version of the charts has been uploaded on OpenAI’s blog post.

Meanwhile, an OpenAI marketing staffer also apologized for the error in a post on X (formerly Twitter), writing, “We fixed the chart in the blog guys, apologies for the unintentional chart crime 🙏”

What’s new with GPT-5?

OpenAI says GPT-5 comes with major improvements in areas like accuracy, speed, reasoning, context recognition, structured thinking, and problem-solving compared to the company’s GPT-4o model.

The biggest change with GPT-5 is the introduction of a unified system with an ‘efficient model’ powering GPT-5’s normal tasks while a dedicated reasoning model called GPT-5 Thinking handles harder reasoning-based tasks.

Unlike in the past, where users had to choose the model for each query, GPT-5 features a real-time router trained on real signals to instantly decide which model to use.

The new model comes with major improvements in coding-related tasks and also possesses the ability to create apps, games and websites using natural language prompts. There is also claimed to be a marked improvement in writing and health-related tasks.