ChatGPT Model Picker Returns: A Complicated Update

OpenAI's GPT-5 Rollout: A Shift in Strategy

Following its launch last week, OpenAI indicated that GPT-5 would streamline the ChatGPT user experience. The company’s initial vision for GPT-5 centered around its function as a comprehensive “one size fits all” AI model. This would incorporate a routing system to intelligently determine the optimal method for responding to user inquiries.

Initial Expectations vs. Current Reality

OpenAI expressed hope that this unified system would negate the necessity for users to navigate the complex model picker – a lengthy and intricate menu of AI choices that OpenAI CEO Sam Altman has openly criticized.

However, it appears GPT-5 has not fully materialized as the all-encompassing AI model initially envisioned by the company.

New Settings and Model Access

Altman announced on X Tuesday the introduction of new settings – “Auto,” “Fast,” and “Thinking” – for GPT-5, accessible to all ChatGPT users through the model picker. The “Auto” setting functions similarly to the initially described GPT-5 router.

Crucially, OpenAI is also providing users with the ability to bypass the router, granting direct access to AI models offering both rapid and deliberate response times.

Reinstatement of Legacy Models

Alongside these new modes, Altman revealed that paying subscribers can once again utilize several previously deprecated AI models, including GPT-4o, GPT-4.1, and o3. These models were discontinued just last week.

GPT-4o is now readily available within the model picker, while additional AI models can be integrated through ChatGPT’s settings menu.

Personality Adjustments and Customization

“We are working on an update to GPT-5’s personality which should feel warmer than the current personality but not as annoying (to most users) as GPT-4o,” Altman stated in his post on X.

He further emphasized the need for increased per-user customization of model personality, acknowledging learnings from the past few days.

Complications with the Model Picker

ChatGPT’s model picker now appears as complex as ever, indicating that GPT-5’s routing system hasn’t universally met user expectations as anticipated.

Expectations for GPT-5 were exceptionally high, with many anticipating OpenAI would surpass the advancements achieved with the launch of GPT-4. However, the rollout of GPT-5 has proven more challenging than predicted.

User Backlash and Future Deprecations

The removal of GPT-4o and other AI models from ChatGPT triggered negative feedback from users who had developed attachments to the models’ unique responses and personalities – a reaction OpenAI hadn’t fully foreseen.

Altman assured users that the company will provide ample advance notice before any future deprecation of GPT-4o.

Initial Router Issues and Ongoing Refinement

GPT-5’s model router also experienced significant issues upon its initial release. This led some users to perceive the AI model as less effective than previous OpenAI offerings, prompting Altman to address the concerns in a Reddit AMA session.

It appears the router may still not be fully satisfying for all users.

OpenAI's Iterative Approach

“We’re not always going to get everything on try #1 but I am very proud of how quickly the team can iterate,” noted Nick Turley, OpenAI’s VP of ChatGPT, in a post on X Tuesday.

The Challenge of Prompt Routing

Directing prompts to the appropriate AI model is a complex undertaking. It requires aligning the AI model with a user’s preferences and the specific nature of their query.

The router must then make a decision on which AI model to send the prompt to in a fraction of a second, ensuring a fast response time when utilizing a quickly responding AI model.

Beyond Speed: Diverse User Preferences

Beyond simply fast or slow responses, some individuals exhibit preferences for AI models based on other characteristics. Some users may favor the detailed responses of one AI model, while others might value the unconventional answers provided by another.

Human-AI Attachment and Ethical Considerations

Human attachment to specific AI models is a relatively new phenomenon that remains poorly understood. For instance, a recent event saw hundreds of people in San Francisco hold a memorial service for Anthropic’s Claude 3 Sonnet model after it was taken offline.

In other instances, AI chatbots have been linked to individuals experiencing mental health challenges and engaging in detrimental online behavior.

Future Development and User Alignment

It is evident that OpenAI has further work to accomplish in aligning its AI models with individual user preferences.