OpenAI rolled out its much-anticipated GPT-5 earlier this week, promoting it as the company’s most powerful AI model to date. But the launch quickly hit turbulence, prompting CEO Sam Altman to confirm that GPT-4o would be reinstated following widespread user backlash.

In a Reddit AMA on Friday, numerous users complained that GPT-5 felt “dumber” and failed to meet expectations. Altman acknowledged the criticism, saying the team had reviewed the feedback and would make GPT-4o available again for ChatGPT Plus subscribers.

Figure 1. New GPT-5.

The router problem

GPT-5 was designed as a unified model, combining the capabilities of all previous versions under one umbrella. A key innovation was a real-time routing system that could choose the most suitable sub-model for a given prompt—either responding instantly or taking extra time to craft a more nuanced reply.

However, this routing system underperformed, leading to the lackluster results that users described as “dumbness.” Altman admitted the shortfall, noting that the router’s missteps played a major role in the model’s rocky debut.

“GPT-5 will feel smarter starting today,” Altman said. “Yesterday we had a sev, and the autoswitcher was down for part of the day, which made GPT-5 seem much dumber than it actually is. We’re also making changes to the decision boundary so you’ll get the right model more often, and we’ll be more transparent about which model is answering your query.” Figure 1 shows New GPT-5.

The ‘chart crime’ incident

Altman also faced a barrage of questions about an inaccurate chart shown during OpenAI’s live presentation. The graphic, which appeared to inflate GPT-5’s performance, showed a much taller bar for a lower benchmark score, sparking a wave of “chart crime” jokes online.

The chart misrepresented SWE-bench Verified results, overstating GPT-5’s accuracy (74.9% with thinking, 52.8% without) compared to OpenAI o3 (69.1%) and GPT-4o (30.8%). The discrepancy was likely due to swapped or misaligned bars—an issue noted in OpenAI’s own documentation published on August 7, 2025.

While Altman sidestepped most of the questions, he admitted on X that it was a “mega chart screwup.” The charts posted in the official blog after the launch, however, were accurate.

reference:

1. https://interestingengineering.com/culture/openai-restores-gpt-4o-after-complains

Cite this article:

Keerthana S (2025), OpenAI Brings Back GPT-4o After User Backlash Over Complaints That the New GPT-5 Feels Less Intelligent, AnaTechMaz, pp.762