Altman on GPT-5 'bumpy' rollout, 4o return, and 'chart crime'
On Friday, OpenAI CEO Sam Altman, alongside key members of the GPT-5 development team, engaged with users in a Reddit ask-me-anything (AMA) session, addressing a range of concerns following the new model’s recent launch. Discussions largely centered on the perceived performance of GPT-5, widespread requests for the return of its predecessor, GPT-4o, and a notable presentation gaffe dubbed the “chart crime.”
Many users voiced frustration, claiming GPT-5 appeared “dumber” or less capable than GPT-4o. Altman acknowledged these complaints, explaining that the issue stemmed from a malfunction in GPT-5’s new real-time router. This innovative feature is designed to dynamically select the most appropriate underlying model for a given query—either providing a rapid response or taking more time to “think” through complex answers. However, on Thursday, the day of the rollout, this crucial routing system experienced a critical service outage, leading to the degraded performance users observed.
Altman assured the community that fixes were already in place. “GPT-5 will seem smarter starting today,” he wrote, confirming that the autoswitcher malfunction had been resolved. He further committed to refining the logic governing model selection to ensure users are more consistently directed to the optimal model. Additionally, OpenAI plans to enhance transparency, making it clearer to users which model is responding to their queries. Responding to persistent appeals, Altman also indicated that OpenAI is exploring the possibility of allowing Plus subscribers to revert to GPT-4o, stating, “We are trying to gather more data on the tradeoffs.” To further ease the transition and encourage exploration of the new model, OpenAI will double the rate limits for Plus users as the rollout concludes, effectively expanding the number of prompts they can make.
Beyond performance concerns, Altman was inevitably questioned about the “chart crime”—a widely mocked, wildly inaccurate chart presented during the GPT-5 live demonstration. The chart, intended to illustrate benchmark scores, misleadingly depicted a lower score with a significantly taller bar, immediately drawing criticism and becoming the subject of countless online jokes. While Altman did not directly address the chart during the AMA, he had previously acknowledged it on X (formerly Twitter) as a “mega chart screwup.” It was later noted that the charts published in the official blog post were corrected. The incident, however, fueled a wave of humorous commentary about using AI for corporate presentations, with one reviewer even highlighting GPT-5’s failure to accurately convert data into a table as an example of its current limitations.
Altman concluded the AMA by reiterating OpenAI’s commitment to improving the user experience. He promised ongoing efforts to stabilize the system and emphasized the company’s dedication to listening to community feedback.