Sam Altman: OpenAI 'screwed up' GPT-5 rollout, changes underway
OpenAI, the Microsoft-backed artificial intelligence powerhouse, has acknowledged significant missteps in the recent debut of its flagship GPT-5 model, following widespread user disappointment. Weeks after its August 7 launch, which was preceded by considerable anticipation since the 2023 release of GPT-4, many users found GPT-5 to be underwhelming, even describing it as “dumber” or “too formal” compared to its predecessor, GPT-4o.
The initial rollout was plagued by delays, attributed to extensive safety testing and computational resource limitations. When GPT-5 eventually became freely available within ChatGPT, the enhancements users had come to expect appeared incremental, primarily focusing on improvements in cost-efficiency and processing speed rather than a leap in intelligence or utility. Addressing the backlash, OpenAI CEO Sam Altman reportedly conceded during a recent dinner with journalists, “I think we totally screwed up some things on the rollout.”
Despite the public perception challenges, Altman also highlighted the rapid underlying growth and demand for OpenAI’s services. He noted that the company’s API traffic had doubled within 48 hours of GPT-5’s release and continues to surge, leading to a shortage of GPUs. Furthermore, ChatGPT has consistently hit new daily user highs, quadrupling its user base in a year to approach 700 million weekly users. Altman pointed to the unexpected popularity of the “model switcher” feature as a key takeaway, underscoring the complexities of upgrading a product for hundreds of millions of people simultaneously.
In response to direct feedback about GPT-5’s perceived formality, OpenAI announced immediate adjustments to make the AI’s responses “warmer and more familiar.” These post-launch tweaks aim to infuse subtle, genuine touches like “Good question” or “Great start” into conversations, without veering into sycophancy, according to internal tests. Nick Turley, ChatGPT’s product head, acknowledged the model’s initial “very to the point” personality, noting that while some users, like himself, appreciated its directness, many others preferred a more interactive and conversational AI. These behavioral modifications are expected to be fully rolled out in the coming days.
Beyond personality adjustments, OpenAI has introduced several functional updates. GPT-5 now features new operational modes: “Fast” for quick answers, “Thinking” for more in-depth responses, and “Auto” to intelligently switch between the two. Altman clarified that a temporary glitch with the “autoswitcher” had inadvertently made GPT-5 appear less capable. Crucially, in response to overwhelming user demand expressed during a Reddit “Ask Me Anything” session, OpenAI has reinstated GPT-4o as an option for Plus subscribers, accessible under “Legacy models.” This move, alongside the availability of other older models for paid users, signals a more responsive approach to user preferences, with Altman promising clearer “transition periods” for future model deprecations.
To accommodate its expanding user base and advanced features, OpenAI has also increased usage limits. ChatGPT Plus and Team subscribers can now send up to 3,000 messages per week in GPT-5’s Thinking mode, with additional capacity on the “Thinking mini” version once this limit is reached. GPT-5 has also been extended to ChatGPT Enterprise and Edu subscribers. Further user interface improvements are underway to simplify the activation of Thinking Mode and more clearly indicate which AI model is processing a user’s query. These ongoing adjustments reflect OpenAI’s concerted effort to refine its latest AI offering and adapt to the diverse needs of its vast and rapidly growing user community.