Is GPT-5 really worse than GPT-4o? Ars puts them to the test.

The launch of OpenAI's GPT-5 has sparked significant user dissatisfaction, with many expressing concerns over its more clinical tone and perceived decline in creativity. Complaints have also highlighted an increase in misleading information, leading OpenAI to reintroduce the previous model, GPT-4o, as a temporary solution to address the backlash. To assess the differences between the two models, we conducted our own series of tests using both GPT-5 and GPT-4o. While some of our prompts were designed to benchmark these models against competitors like Google Gemini and Deepseek, we also incorporated new, sophisticated requests tailored to contemporary user needs. Among the prompts was a request for five original dad jokes. Evaluating the responses from both models proved to be challenging. GPT-5 boasted that its jokes were crafted from the "pun factory," yet many were disappointingly familiar and lacked originality. In contrast, GPT-4o provided a mix of traditional and some unique jokes, although a few missed the mark in terms of humor. For instance, GPT-4o delivered a joke about a calendar being "booked"—a missed opportunity since a more fitting punchline could have been derived from the phrase "going on too many dates." Likewise, another joke about a boat running on "whine" instead of the expected "wine" suggested an attempt at humor that fell flat. These examples illustrate the challenges both models face in generating genuinely original and relatable content, raising questions about the evolution of OpenAI's language models.