GPT-5 launch stumbles prompt swift return of GPT-4o

OpenAI introduced GPT-5 as its new unified flagship, positioning it as "the most powerful model yet." Within a day, however, CEO Sam Altman acknowledged a rocky debut and confirmed that GPT-4o will be reinstated as a selectable option for ChatGPT Plus subscribers while the company evaluates usage. The rapid reversal followed a wave of complaints from users who felt the update removed a preferred experience without warning.

GPT-5 consolidates prior models behind a real-time routing system intended to pick the best engine for each prompt, sometimes answering immediately, sometimes "thinking" longer for tougher queries. Altman said an outage affecting the autoswitcher contributed to perceptions that GPT-5 was less capable, and he pledged adjustments to the decision boundary and greater transparency about which model responds to a given request.

Beyond performance concerns, the launch reignited debate over AI "personality." Many long-time users associated GPT-4o with a warmer, more conversational tone and described GPT-5 as colder and more mechanical, even if technically strong. Threads on Reddit and posts on X captured frustration with the removal of model choice and, in some cases, vows to cancel subscriptions if GPT-4o did not return. The company now says GPT-4o's long-term availability will depend on actual demand.

The rollout was also marred by a charting error during launch. A bar chart referencing SWE-bench Verified appeared to misrepresent benchmark results, showing much taller bars for lower scores, an issue Altman later called a "mega chart screwup." OpenAI's blog was updated with corrected figures, and separate documentation dated August 7, 2025, reflects the proper accuracy numbers cited in the discussion.

Altman fielded extensive questions in a Reddit AMA, reiterating that GPT-5 would "seem smarter" as the routing system stabilized and promising to "keep listening to feedback." He also said the company is "looking into letting Plus users continue to use 4o" while it gathers more data on trade-offs between models. For now, OpenAI plans to monitor usage before deciding how long to support GPT-4o alongside GPT-5.