Key Takeaways
- GPT-5 launched with the goal of doing away with ChatGPT’s model picker.
- User pushback and system malfunctions have brought the picker back.
- New modes “Auto,” “Fast,” “Thinking,” and legacy models have been added.
- Paid users can access GPT-4o, GPT-4.1, and o3 again.
- OpenAI’s future plans include adding more customizable personalities for the models.
GPT-5’s “One-Model Strategy” Falls Short
Last week when OpenAI launched GPT-5, it promised a more streamlined ChatGPT interface.
It focused on one model having a built-in router, which would make autonomous decisions on how to respond to the prompt.
- Thereby, eliminating the need of the model picker.
- But just days later, the picker is back.
- Old Models and New Modes Come Back
- On Tuesday, Sam announced three additional modes for GPT-5:
Auto: allows the router to choose the AI model.
Fast: prioritizes speed.
Thinking: detailed slow reasoning.
Paid users can also re-enable GPT-4o, GPT-4.1, and o3 from ChatGPT’s Model Picker.
GPT-4o is on by default but the other models need to be added.
Altman also provided specifics on how he is improving GPT-5’s persona. He says it will be “warmer” but less “annoying” relative to GPT-4o.
Why Users Pushed Back
- To start, OpenAI’s removal of older models last week, infuriated a lot of users.
- Some users were fond of a given models slowness or style, and the way it ‘interacted’ and its persona.
- This attachment is stronger than many (including the authors) expected.
- For instance, in San Francisco, hundreds showed up to a mock memorial for Claude 3 Sonnet, an Anthropic model, when it got discontinued.
- Thanks to user feedback, Altman said future model removal policies (such as the expected deprecation of GPT-4o) will be signaled earlier.
Technical and Preference Challenges
On GPT-5’s launch day, the model router’s performance for ChatGPT’s Model Picker users to complain.
As of today, it appears that a fair share of users are still in disagreement on how the model selection is conducted.
- Routing is a complex endeavor.
- It should balance user expectations with the type of question being answered.
- It has to select a model in a few milliseconds, or else performance will take a hit.
However, a model selection prioritizing speed comes with its own problems. Some users have preferences for curt responses and others for longer answer.
Meeting the expectation of all of these is likely to need more customization user-side.
What’s Next for ChatGPT’s Model Picker
- OpenAI leadership says that the work on GPT-5 is under a ‘Get it right, get it fast’ strategy.
- Nick Turley says the company understands it will not “get everything right on try #1” but promises swift iterations.
- Since user feedback is an integral part of product decision-making, it’s possible that the model picker could return for good — even in the GPT-5 era.
Synopsis
- GPT-5 intended to switch the ChatGPT’s Model Picker for an AI router.
- Consumer pushback coupled with technical issues saw the return of the picker.
- There are preset new modes (“Auto,” “Fast,” “Thinking”) alongside legacy models.
- OpenAI is looking to implement comprehensive model-specific user instructions for personality customization and model communication pre-removal.
- In the future, the user-centric experience may center on personalized settings instead of default configurations.