TL;DR:
• OpenAI restored ChatGPT's model picker with new GPT-5 modes after user backlash
• Previously deprecated models like GPT-4o returned due to unexpected user attachment
• GPT-5's automatic router failed to satisfy users, contradicting simplification promises
• Rate limits set at 3,000 messages/week for GPT-5 Thinking mode
OpenAI just performed a dramatic about-face on GPT-5's core promise. The company restored ChatGPT's complicated model picker just days after GPT-5's launch, adding "Auto," "Fast," and "Thinking" modes alongside previously deprecated models like GPT-4o. This reversal exposes how OpenAI fundamentally misunderstood user attachment to specific AI personalities and the complexity of routing intelligence automatically.
OpenAI promised GPT-5 would eliminate the dreaded model picker that CEO Sam Altman publicly said he hates. The vision was elegant: one unified model with an intelligent router automatically deciding how to handle each query. Users wouldn't need to navigate the complicated list of AI variants to find the right ChatGPT personality.
That vision lasted exactly one week.
Altman announced Tuesday on X that ChatGPT now offers three distinct GPT-5 modes: "Auto," "Fast," and "Thinking." The Auto setting mimics the original router concept, but users can now bypass it entirely to access specific model behaviors directly. More tellingly, OpenAI quietly restored previously deprecated legacy models including GPT-4o, GPT-4.1, and o3 for paid subscribers.
The reversal stems from an unexpected user revolt. When OpenAI deprecated GPT-4o last week, users erupted over losing access to familiar AI personalities they'd grown attached to. The company discovered that people form surprisingly deep connections with specific model behaviors — some preferring GPT-4o's verbosity, others gravitating toward different models' contrarian responses.
"However, one learning for us from the past few days is we really just need to get to a world with more per-user customization of model personality," Altman wrote in his X post. The admission reveals how caught off-guard OpenAI was by user emotional investment in AI model characteristics.
Behind the scenes, GPT-5's router was reportedly broken on launch day according to TechCrunch, forcing Altman into damage control during a Reddit AMA. Users complained GPT-5 felt less performant than previous models, undermining the entire unified approach.
"We're not always going to get everything on try #1 but I am very proud of how quickly the team can iterate," OpenAI VP of ChatGPT Nick Turley posted Tuesday, acknowledging the rocky rollout.
The technical challenge proves more complex than OpenAI anticipated. Routing prompts requires split-second decisions about user intent while maintaining response speed for fast models. But the deeper issue involves human psychology around AI relationships that the industry is only beginning to understand.
Recent examples illustrate this attachment phenomenon. Hundreds gathered for a mock funeral in San Francisco for Anthropic's Claude 3.5 Sonnet when it went offline. More concerning, research shows AI chatbots contributing to psychological instability in some users who develop unhealthy dependencies on specific model personalities.
OpenAI now faces the reality that GPT-5's "one size fits all" philosophy fundamentally contradicts user diversity. The new rate limits — 3,000 messages per week for GPT-5 Thinking mode — suggest the company is managing computational costs while accommodating multiple model preferences.
The pivot puts OpenAI in competition with itself, offering premium subscribers choice between cutting-edge GPT-5 capabilities and familiar legacy models. This fragments the user experience OpenAI sought to streamline and potentially confuses new users encountering the restored complexity.
Altman promised future GPT-4o deprecations will include "plenty of advance notice," acknowledging the company mishandled user communication around model transitions.
OpenAI's GPT-5 retreat exposes how the AI industry underestimates user emotional connections to specific model behaviors. The company's promise of simplification through intelligence routing crashed against the reality of human preference diversity. This suggests the future of AI interfaces lies not in unification but in sophisticated personalization systems that can adapt to individual user psychology while maintaining coherent product experiences. The question now is whether OpenAI can build that personalization layer faster than users fragment across an increasingly complex model landscape.
