AI Issue: GPT-5 Release Issue -- What Comes Next?
Sam Altman Addresses GPT-5 Launch Issues and Relaunches GPT-4o for Users
Key Highlights
- On GPT-5's launch day, the router model malfunctioned, causing performance issues.
- Sam Altman announced that GPT-4o and other legacy models would be made available again for ChatGPT users.
- OpenAI added a model selector with auto, fast, and thinking modes along with legacy models.
- Altman promised to double usage limits for Plus subscribers during the GPT-5 rollout.
- They are exploring ways to customize model personality differently for each user.
Problems That Emerged During GPT-5's Launch
When OpenAI released GPT-5 last week, it was introduced as a "one-size-fits-all" solution.
This unified approach was supposed to eliminate the need for users to manually select models, as OpenAI CEO Sam Altman had publicly expressed his dislike of long, complex AI option menus.
The AI was designed with a real-time router that automatically determines whether to answer quickly or think more deeply, eliminating the need for a long model selection menu.
However, the router malfunctioned on launch day, leading some Reddit users to rate GPT-5 as performing worse than the previous version, GPT-4o.
In a Reddit AMA on Friday, CEO Sam Altman explained:
"GPT-5 is going to look smarter starting today. Yesterday there was a really big issue where the auto-switcher wasn't working all day, which made GPT-5 look much dumber."
He added that OpenAI is adjusting "the way it decides" so users get the right model more often, and will make it clearer which model is answering.
The Return of Model Selection
A few days after the launch, Altman announced on X that ChatGPT users can now once again select from legacy AI models alongside the auto, fast, and thinking modes in the model selector.
For paid subscribers, GPT-4o, GPT-4.1, and o3 are once again available.
GPT-4o appears by default in the selector, and other models can be added from settings.
Altman accepted users' frustration about suddenly losing GPT-4o, saying he would "give plenty of notice" when models will no longer be available in the future.
User Reactions and Feature Improvements
During the AMA, users argued that GPT-4o worked better for them and requested its return.
Altman responded: "We're finding a way for Plus users to keep using 4o. We're trying to gather more information about the pros and cons."
He also promised to double the usage limits for Plus users, giving them more opportunity to use GPT-5 without hitting restrictions.
Along with the feature changes, Altman revealed that work is underway to make GPT-5's personality warmer.
"We're trying to update GPT-5 to a personality that feels warmer than the current one but not as annoying as GPT-4o (for most users)."
"But one thing we've learned over the past few days is that we need a world where model personality can be more personalized for each user," Altman wrote on X.
The "Chart Fail" Incident
The GPT-5 launch event also attracted attention for an embarrassing graph error dubbed the "chart fail."
A presentation slide showed low benchmark scores with disproportionately large bar charts, becoming a target of mockery on social media.
While Altman didn't directly address the issue during the AMA, he had called it an "epic chart fail" on X the day before.
The blog post contained the correct charts, but the incident became part of the early GPT-5 narrative amid mild criticism.
The Difficult Challenge of Model Calibration
OpenAI executives acknowledge that the routing approach is still complex.
Nick Turley, OpenAI's VP of ChatGPT, said on X: "You can't always get everything perfect on the first try, but I'm really proud of how quickly the team can iterate."
Routing decisions need to balance user preferences, query types, and model capabilities, and this typically needs to happen within milliseconds.
But user attachment to specific models goes beyond speed or accuracy -- some prefer a particular style or personality.
This attachment has become a cultural phenomenon: TechCrunch reported that hundreds of people in San Francisco held a funeral after Anthropic's Claude 3.5 Sonnet was discontinued.
Altman concluded the AMA saying, "We'll continue to work on stabilization and keep listening to feedback."
Q&A
Q: Why did GPT-5 look 'dumb' at launch?
A: The router model malfunctioned on launch day, causing performance issues.
Q: What changes happened at OpenAI afterward?
A: GPT-4o and other legacy models were made available again, the model selector was re-added, and auto, fast, and thinking modes were introduced.
Q: Do Plus users get additional benefits?
A: Yes. Usage limits for Plus subscribers were doubled during the rollout period.
Q: Is GPT-5's personality changing?
A: Yes, OpenAI is working to create a warmer personality and exploring ways to customize it per user.
Q: What is the 'chart fail'?
A: A graph error in the launch presentation that showed incorrect bar chart proportions, later corrected in the blog post.
Implications
The GPT-5 launch demonstrates both the promise and challenges of consolidating AI capabilities into a single model.
The router system aims to simplify the user experience, but early issues and strong attachment to older models like GPT-4o show that technical efficiency alone isn't the only factor in user satisfaction.
By bringing back legacy models and adding personality customization, OpenAI is signaling that it values user trust and flexibility alongside technological advancement.
While GPT-5's early problems may slow its progress toward a simplified AI experience, they also provide lessons on balancing the new with the familiar in the fast-moving world of AI.
IMPAKERS Blog | New OpenAI Model Coming? Read More
Source: Alicia Shapiro, AiNews, "Sam Altman Addresses GPT-5 Rollout Issues, Brings Back GPT-4o for Users", https://www.ainews.com/p/sam-altman-addresses-gpt-5-rollout-issues-brings-back-gpt-4o-for-users, (2025. 8. 13)