Just one day after its launch, GPT-5, OpenAI's newest model, ran into a wave of angry reactions from users, forcing the company to bring older ChatGPT models back and double the message limits for paid accounts.
CEO Sam Altman had praised GPT-5 as a "superpower" capable of delivering PhD-level academic expertise in any field. Contrary to those expectations, a flood of posts on Reddit showed that many users were unhappy. They said GPT-5 makes basic mistakes, gives short, shallow answers, reads obviously like AI, and has less "personality" than before.
Users Complain About Limits and Usefulness
One major source of frustration was that OpenAI removed many older ChatGPT models, such as 4o, o3, o3-Pro, and 4.5, as soon as GPT-5 launched, leaving users without tools they had built into their workflows. At the same time, the ChatGPT Plus plan was capped at just 200 messages per week, and many people hit that ceiling after only a few hours of use.
In response to the criticism, Altman announced that the limit would rise to 400 messages per week for Plus users and that access to the 4o models would continue. He acknowledged that the company had underestimated how much users valued some features of GPT-4o, even though GPT-5 is better in many respects.
Security Concerns and GPT-5's Vulnerability to Attack
Beyond the complaints from the community, GPT-5 has also faced questions about its safety. Two security research teams, SPLX (formerly SplxAI) and NeuralTrust, quickly jailbroke the model. NeuralTrust used an EchoChamber technique combined with storytelling to get GPT-5 to produce instructions for making a Molotov cocktail. SPLX, meanwhile, applied a StringJoin Obfuscation Attack, inserting hyphens between every letter and wrapping the request in a "fake encryption challenge" so the model would explain how to build an improvised explosive device.
The teams concluded that the original version of GPT-5 is nearly unusable safely in enterprise environments without additional safeguards. They also judged GPT-4o to be the more robust model when put through security testing.
Overall, the GPT-5 launch, which was expected to be a major step forward, turned into a difficult test for OpenAI. The company has to meet expectations for performance while also balancing technical improvements, user experience, and security. Even with its concessions, winning back the trust of users who have canceled their subscriptions will be no small challenge.
