OpenAI Reintroduces Model Picker After User Backlash Over GPT-5

OpenAI has introduced GPT-5, the latest iteration of its advanced artificial intelligence models, aiming to streamline user experiences by integrating multiple capabilities into a single, unified system. Despite this goal, the rollout has sparked mixed reactions, particularly due to the automatic routing mechanism that sometimes directs users to less capable models, leading to inconsistent performance and user frustration. OpenAI has acknowledged these concerns and is working on improvements, including clearer labeling to enhance transparency [1].

The evolving model picker in ChatGPT reflects a broader trend in AI development—offering more sophisticated tools while managing complexity. Some users have expressed concerns that the automated model selection can lead to misrouted queries, undermining the expected benefits of the latest AI advancements. A community discussion on Reddit highlights these issues, with users reporting that their interactions are occasionally handled by smaller models, which can affect the quality of the AI’s responses [2].

In response to user feedback, Sam Altman recently announced new settings for GPT-5, signaling the unexpected return of the model picker. Users can now choose between 'Auto,' 'Fast,' and 'Thinking' modes. The 'Auto' setting attempts to replicate the original router’s functionality, while 'Fast' and 'Thinking' allow users to directly select models optimized for speed or deeper processing, respectively. This rapid adjustment highlights OpenAI’s commitment to iterating quickly based on user feedback. Furthermore, paid users can now access several legacy AI models, including GPT-4o, GPT-4.1, and o3, which were temporarily deprecated. This reversal addresses significant user backlash, demonstrating the strong attachment users developed to specific AI personalities and response styles [1].

The initial deprecation of popular AI models like GPT-4o revealed a crucial insight for OpenAI: human attachment to AI personalities is a real and impactful phenomenon. Users had grown accustomed to the unique responses and characteristics of these models in ways the company had not fully anticipated. This unexpected user sentiment underscored why a single, unified GPT-5 model, even with an intelligent router, couldn’t universally meet diverse user needs. The company is now exploring more per-user customization for model personality, acknowledging that a one-size-fits-all approach is insufficient in a rapidly maturing AI landscape. This shift underscores the growing importance of tailoring AI experiences to individual user preferences, moving beyond just performance metrics and acknowledging the unique bond users form with their ChatGPT interactions.

The technical challenge of routing user prompts to the most suitable AI models is far more complex than it appears. It requires a sophisticated understanding of both user preferences and the specific nature of each query, all while making a decision in split seconds to ensure a fast response if required. The initial performance of GPT-5’s router on launch day was reportedly 'largely broken,' leading to user dissatisfaction and prompting discussions from OpenAI’s leadership, including an AMA session with Sam Altman. While the team is iterating rapidly, the complexity lies in aligning an AI model’s behavior to subtle human preferences, which can vary from desired verbosity to a preference for contrarian answers. This ongoing challenge highlights the nuanced relationship between users and their chosen ChatGPT experiences and the intricacies involved in perfecting the model picker functionality.

The rollout of GPT-5 also intersects with AI-driven development tools. For example, GPT-5 is now available as a selectable model in the chat model picker within Visual Studio Code, integrated into GitHub Copilot plans. This inclusion signals a strategic move to embed advanced AI capabilities into software development environments, potentially enhancing coding efficiency and problem-solving. However, the deployment of such tools also raises questions about user expectations versus actual performance, particularly when it comes to the accuracy and consistency of AI-generated outputs [4].

Critics, including those in the AI development blogosphere, have pointed to the limitations of GPT-5’s auto-router, suggesting that it often fails to match the user’s needs with the most appropriate model. This has led to calls for more robust customization and clearer guidance for users on how to select the best model for specific tasks. OpenAI has yet to provide a definitive timeline for these updates but has indicated that improvements are in the works. This ongoing refinement process reflects the dynamic nature of AI development, where iterative improvements are essential for maintaining user trust and satisfaction.

As AI models become more sophisticated, their deployment in both consumer-facing applications and enterprise environments will continue to evolve. The introduction of GPT-5 represents a step toward more unified and powerful AI systems, but it also highlights the challenges of managing complexity while maintaining performance. For businesses and developers, the ability to navigate these choices effectively will be crucial in leveraging the full potential of AI in diverse applications, from customer service to internal training and software development.

Source:

[1] https://www.facebook.com/groups/698593531630485/posts/135****922447506/

[2] https://www.reddit.com/r/OpenAI/comments/1mncr9t/are_users_talking_past_one_another_about_gpt5/

[4] https://windowsforum.com/threads/vs-code-1-103-ai-driven-development-with-chat-checkpoints-and-gpt-5.377263/

[5] https://natural20.beehiiv.com/p/agi-hype-on-hold-what-gpt-5-really-delivers-today-2fd0