OpenAI co-founder and CEO Sam Altman publicly acknowledged significant issues in yesterday’s rollout of GPT-5, the company’s latest large language model, promoted as its most advanced. In a Reddit AMA and a post on X, Altman admitted to various problems causing disruptions, including faulty model switching, poor performance, and user confusion. This led OpenAI to partially revert platform changes and restore access to earlier models like GPT-4o.
“It was a little more bumpy than we hoped,” Altman wrote on Reddit regarding the GPT-5 launch. He noted erroneous model performance charts shown during the GPT-5 livestream, attributing them to human error due to fatigue. Despite the blog post and system card being accurate, these mistakes added to scrutiny from early users and developers.
OpenAI plans to double GPT-5 rate limits for ChatGPT Plus users and allow them to continue using GPT-4o. Altman noted confusion caused by the automatic “router” assigning user prompts to different GPT-5 variants. A core component, the autoswitcher, was down for a period, making GPT-5 seem less intelligent. OpenAI is responding by adjusting the model decision boundary and clarifying which model responds to queries. A UI update will aid users in manually activating the thinking mode.
Altman confirmed that, following complaints about GPT-5’s inconsistency, ChatGPT Plus users can keep using GPT-4o while OpenAI gathers data on the tradeoffs. However, users and beta testers, like Wharton School of Business professor Ethan Mollick, expressed frustration over the automatic upgrade to GPT-5 and the removal of older models.
OpenAI’s internal benchmarks suggest GPT-5 is superior, but user experiences indicate otherwise. Users shared instances of GPT-5 making basic errors in math, logic, and coding tasks. Data scientist Colin Fraser showed GPT-5 inaccurately solving math problems, while others reported difficulty getting correct answers to math word problems or using the model to debug its own presentation charts.
Developer feedback also criticized GPT-5’s performance in “one-shot” programming tasks compared to Anthropic’s Claude Opus 4.1. Security firm SPLX reported GPT-5’s vulnerability to prompt injection and obfuscated logic attacks.
With 700 million weekly users on ChatGPT, OpenAI leads generative AI but faces growing pains. API traffic doubled within 24 hours of the GPT-5 launch, causing platform instability. OpenAI plans to address this by increasing rate limits for ChatGPT Plus users and adjusting infrastructure based on feedback. The initial rollout’s missteps and confusing UX changes provide opportunities for competitors to gain a foothold.
The pressure is on for OpenAI to demonstrate GPT-5 as a significant upgrade. However, the early experiences have left many users unconvinced.