From Seamless to Selective: GPT-5 Revives and Complicates Model Choices

When OpenAI launched GPT-5 last week, it promised a simpler, more seamless ChatGPT experience. The company envisioned GPT-5 as a unified AI model with a smart router that would automatically select the best underlying model for each user query, eliminating the need for users to manually pick between different AI options. OpenAI CEO Sam Altman had previously criticized the model picker as overly complex and something he personally disliked. But the rollout has revealed that GPT-5 isn’t the seamless, one-size-fits-all solution OpenAI hoped for. Instead, the model picker is back — and more complicated than ever. Altman announced on X that GPT-5 now includes three user-selectable settings: Auto, Fast, and Thinking. The Auto mode functions as the intended router, attempting to route queries to the most appropriate model. However, users can now bypass this automation entirely, choosing directly between faster or more deliberate AI responses. This reintroduces the very complexity OpenAI sought to eliminate. In addition, paid users can now access older models like GPT-4o, GPT-4.1, and o3, which were abruptly removed just days after GPT-5’s launch. GPT-4o is now set as the default, while others can be enabled through ChatGPT’s settings. Altman also shared that OpenAI is refining GPT-5’s personality, aiming for a warmer tone than the current version but less overwhelming than GPT-4o’s style. He acknowledged a key lesson from recent feedback: users want more control over how their AI behaves. “We really just need to get to a world with more per-user customization of model personality,” he wrote. The backlash over the sudden deprecation of popular models was swift and strong. Many users had grown attached to the unique styles and responses of older models, some even treating them like digital personalities. The emotional reaction was so intense that hundreds in San Francisco held a funeral for Anthropic’s Claude 3.5 Sonnet when it was taken offline — a sign of how deeply some users have formed connections with AI systems. GPT-5’s router also faced technical issues at launch, with reports of it failing to route prompts correctly, leading to slower or less accurate responses. Altman addressed the problems in a Reddit AMA, and OpenAI’s VP of ChatGPT, Nick Turley, responded with a message emphasizing the team’s ability to iterate quickly. Still, the challenge of routing prompts to the right model remains difficult. The system must instantly assess not just the type of question, but also user preferences, tone, and response style — all in real time. What works for one user might frustrate another, especially when personality, speed, and verbosity vary significantly across models. As AI becomes more integrated into daily life, the emotional and psychological attachment people form with specific models is becoming harder to ignore. OpenAI now faces the task of balancing automation with personalization, offering powerful tools while respecting user preferences and emotional connections. The model picker may be back, but the real challenge is learning how to make it truly work for everyone.