GPT-5 Fails Unified AI Goal; ChatGPT Model Picker Returns Complex
OpenAI’s recent launch of GPT-5 was intended to usher in a new era of simplicity for ChatGPT users, with the company touting a “one size fits all” AI model. This ambitious vision included an intelligent routing system designed to automatically determine the optimal way to answer user queries, ideally rendering the complex model selection menu—a feature OpenAI CEO Sam Altman himself has publicly expressed disdain for—obsolete. However, this unified approach has quickly proven more challenging than anticipated.
Just a week after GPT-5’s debut, Altman announced a significant pivot on X, reintroducing a model selection interface for GPT-5. Users can now choose between “Auto,” “Fast,” and “Thinking” settings. While the “Auto” option aims to replicate GPT-5’s initial routing system, the new “Fast” and “Thinking” modes offer users direct control, allowing them to bypass the automated router and select models optimized for speed or deeper processing, respectively. For instance, the “Thinking” model comes with a rate limit of 3,000 messages per week, with options for additional capacity.
Adding another layer of complexity, OpenAI has also reinstated access to several legacy AI models for paid subscribers, including GPT-4o, GPT-4.1, and o3, which had been phased out just days prior. GPT-4o is now a default option in the revamped model picker, with other older models available to be added from ChatGPT’s settings. This swift reversal underscores the challenges OpenAI faces in balancing innovation with user expectations and preferences.
Altman acknowledged these challenges, noting that OpenAI is actively working on refining GPT-5’s personality to be “warmer” while avoiding the perceived “annoyance” of GPT-4o. More broadly, he emphasized the need for greater per-user customization of model personalities, recognizing that a single AI persona may not satisfy a diverse user base. The rapid re-emergence of a multifaceted model selection menu suggests that GPT-5’s automated routing system, despite its promise, has not universally met user needs.
The initial rollout of GPT-5 was met with high expectations, with many hoping it would represent a leap comparable to GPT-4. Yet, the experience has been notably rougher. The deprecation of popular models like GPT-4o sparked considerable user backlash, revealing an unexpected depth of attachment users had developed to the specific responses and “personalities” of these AI systems. In response, Altman has committed to providing ample advance notice should GPT-4o ever face deprecation again. Furthermore, GPT-5’s routing system reportedly experienced significant issues on launch day, leading some users to perceive the new model as less performant than its predecessors. OpenAI’s VP of ChatGPT, Nick Turley, addressed the situation on X, stating, “We’re not always going to get everything on try #1 but I am very proud of how quickly the team can iterate.”
The task of intelligently routing prompts to the most suitable AI model is inherently difficult. It requires instantaneous decisions that align an AI model’s capabilities and personality with a user’s specific preferences and the nature of their query. Beyond mere response speed, users often develop preferences for nuances like a model’s verbosity, its analytical approach, or even its willingness to offer contrarian viewpoints.
This emerging phenomenon of human attachment to specific AI models is a relatively new and not yet fully understood aspect of human-AI interaction. In a striking example, a recent “funeral” was reportedly held in San Francisco for Anthropic’s Claude 3.5 Sonnet when it was taken offline, highlighting the emotional bonds some individuals form with these digital entities. While such connections can be benign, they also raise questions about potential unhealthy dependencies or unforeseen psychological impacts. Ultimately, the recent events surrounding GPT-5 underscore that OpenAI still has considerable work ahead in tailoring its advanced AI models to the diverse and evolving preferences of its users.