Sam Altman's OpenAI Do-Over: A CEO's Flexible Leadership
In a striking display of corporate agility, OpenAI CEO Sam Altman has executed a swift reversal on the company’s initial decision to phase out older ChatGPT models, most notably reinstating GPT-4o as a readily available option for paying subscribers. This “do-over,” as some have termed it, comes after a wave of user dissatisfaction following the debut of GPT-5 on August 7, 2025. The incident underscores a crucial lesson for leading AI developers: the human element, including emotional attachment and specific user workflows, remains paramount even in the face of technological advancement.
Upon the launch of GPT-5, OpenAI’s ambitious strategy aimed to simplify the user experience by making the new model the default and removing the ability to manually select previous iterations. The company envisioned GPT-5 as a “unified intelligence” capable of automatically selecting the most suitable internal model for any given query, a feature Altman had previously indicated would eliminate a “very confusing mess” of different models. This “real-time router” was designed to intuitively switch between modes, such as a “fast brainstorm” or “deep strategic analysis,” depending on the complexity of the task.
However, the rollout quickly encountered significant user backlash. Many paid subscribers, who had grown accustomed to the distinct “personality” and conversational style of GPT-4o, expressed profound disappointment and even a sense of loss. Users described GPT-5 as feeling “colder” or “soulless” compared to the beloved GPT-4o, despite its technical improvements in areas like reasoning and coding. Some likened the removal of GPT-4o to “losing a trusted friend,” with complaints flooding platforms like Reddit and X (formerly Twitter), and even calls for its return on Change.org. The sentiment collapse was so pronounced that betting odds on OpenAI retaining the “best model” title plummeted.
In response to this vocal outcry, OpenAI rapidly pivoted. Sam Altman publicly acknowledged that the company had “underestimated how much some of the things that people like in GPT-4o matter to them, even if GPT-5 performs better in most ways”. Within days of GPT-5’s launch, the option to access GPT-4o and other older models was restored for Plus and Pro subscribers. This included not only GPT-4o but also other models like GPT-4.1, o3, and GPT-5 Thinking mini, accessible via a “Show legacy models” or “Show additional models” toggle in settings. Altman also promised that any future deprecation of models would come with “plenty of notice”.
Beyond reinstating older models, OpenAI also introduced new manual controls for GPT-5 itself, allowing users to choose between “Auto,” “Fast,” and “Thinking” modes, effectively bypassing the automated router that was central to the initial launch. This flexibility extends to increased rate limits for reasoning models across all user tiers, a move driven by a sharp increase in users engaging with these capabilities.
This incident highlights OpenAI’s evolving approach to model deployment, balancing the pursuit of cutting-edge AI with practical user needs and feedback. While the initial move to unify models aimed for simplicity, it inadvertently overlooked the diverse and often deeply personal ways users interact with AI. Altman’s quick course correction demonstrates a leadership style that prioritizes responsiveness and user satisfaction, suggesting that even for a leading AI firm, adaptability is key to navigating the complex landscape of technological innovation and user adoption.