When OpenAI launched GPT-5 last week, it was introduced as an all-in-one solution — a single model with an integrated router designed to automatically select the best reasoning style for any query.
Highlights
- Model Picker Restored: ChatGPT now offers three GPT-5 modes — Auto, Fast, and Thinking — plus the return of older models like GPT-4o, GPT-4.1, and o3.
- User Feedback Drove the Change: Backlash over GPT-4o’s removal and attachment to older model personalities prompted OpenAI to restore manual selection.
- Personality Differences: CEO Sam Altman admitted GPT-5 currently feels “cooler” than GPT-4o and promised warmer, more engaging updates.
- Launch Glitches: GPT-5’s auto-router had early performance issues, leading to weaker outputs until improvements are made.
- Personalization Matters: Users value specific answer styles — from concise to deeply reasoned — making model preference highly individual.
- Multi-Model Under the Hood: GPT-5’s “single model” is actually a router between multiple sub-models like Chat, Thinking, and Thinking Pro.
- Future of Steerable AI: OpenAI is testing customizable personalities such as Listener, Robot, and Nerd to adapt to user preferences in real time.
CEO Sam Altman framed it as a way to eliminate the need for manual model selection, a feature he has publicly said he dislikes.
Only days later, the model picker has returned to ChatGPT. It now offers three GPT-5 options: Auto (router-enabled), Fast (optimized for speed), and Thinking (optimized for deeper reasoning).
The reappearance comes alongside the restoration of several previous models, including GPT-4o, GPT-4.1, and o3. GPT-4o is again available by default, while the others can be re-enabled in settings.
User Backlash Influences the Rollback
The change follows a wave of user dissatisfaction after certain models — especially GPT-4o — were removed. Many users expressed strong attachment to the personalities and output styles of older models, a factor OpenAI appears to have underestimated.
Altman acknowledged that GPT-5’s personality currently feels “cooler” than GPT-4o’s and said updates will make it “warmer” without being polarizing.
Technical Challenges at Launch
Part of the issue was technical. At launch, GPT-5’s router experienced glitches, leading to weaker and less responsive outputs. OpenAI’s VP of ChatGPT, Nick Turley, explained that real-time routing is complex and that the team is iterating quickly to improve it.
Why Personalization Matters
The return of manual selection reflects a deeper challenge: model preference is highly personal. Some users prefer concise answers, others favor detailed reasoning, and some value a model’s tendency to challenge their assumptions.
This attachment can be strong — earlier this year, hundreds in San Francisco held a symbolic “funeral” for Anthropic’s retired Claude 3.5 Sonnet.
GPT-5’s Multi-Model Structure
Although marketed as a single model, GPT-5 is technically a multi-model system. The auto-router dynamically chooses between sub-models — including Chat, Thinking, and Thinking Pro — based on prompt complexity and user behavior patterns.
Steerable AI
OpenAI is exploring greater per-user customization through what it calls “steerability research.” Beyond the standard “Default” tone, experimental personalities such as Listener, Robot, Nerd, and others are being tested.
This aligns with trends in AI research toward reward-factorized personalization, where systems adapt to individual preferences in real time.
OpenAI’s ongoing challenge is to build not just smarter AI, but AI that feels personal enough to retain user trust and engagement.
