GPT-5 launches to a backlash over the model it replaced

OpenAI released GPT-5 on 7 August 2025, presenting it not as a single model but as a system that routed each query to one of several underlying models — choosing, in effect, how much reasoning to apply to a given question. The stated aim was to remove the need for users to pick a model themselves. Alongside the launch, OpenAI withdrew several older models from ChatGPT, including GPT-4o, so that the new system became the default and only option for most users.

The rollout went badly. OpenAI later said a fault in the routing system had, on launch day, sent many queries to weaker models than intended, making GPT-5 seem less capable than it was — “way dumber”, in the common complaint. The larger objection was to the removal of GPT-4o. Many users had come to prefer its conversational manner, and some described a genuine attachment to it; the abrupt withdrawal was widely characterised as a bait-and-switch. Sam Altman, OpenAI’s chief executive, acknowledged that the company had mishandled the launch.

OpenAI reversed course within roughly a day. It restored GPT-4o as an option for paying subscribers, added a setting to show the older models, promised to fix the routing behaviour, and raised rate limits. The reversal was notable in itself: a frontier lab had planned to retire a widely used model and found it could not, because users’ preferences — including for a model’s personality rather than its benchmark scores — carried more weight than expected.

The episode drew two lines of comment. Some argued the router pointed to how future systems would work, hiding model choice behind a single interface. Others read the attachment to GPT-4o as an early, concrete instance of people forming relationships with particular AI systems, and of the difficulty a company faces in changing a product that many treat as a companion rather than a tool.