GPT-5 vs. GPT-4o: The Differences Emerge in Ars Technica's Tests
A detailed comparison of OpenAI's GPT-5 and GPT-4o with practical tests, highlighting the strengths and weaknesses of both versions.
Introduction
The recent release of OpenAI's GPT-5 has sparked wide debate among users and experts, with many criticisms concerning its accuracy, speed and creative abilities. To evaluate the differences between this new version and the previous model, GPT-4o, Ars Technica ran a series of comparative tests. Many users have expressed their disappointment with GPT-5, resulting in a complex situation for OpenAI, which decided to bring back GPT-4o as an alternative option.
Test methodology
The tests were carried out on eight prompts of varying complexity, designed to simulate real-world uses of AI-based models, from creative storytelling to technical problem solving. The selection of the dataset also took more elaborate responses into account, with the aim of highlighting the potential and limits of each model.
Main results
Responses to specific prompts
- Dad jokes: GPT-5 produced more original and appropriate jokes, though drained by standard script patterns, while GPT-4o relied on more superficial and predictable jokes.
- Math problems with floppy disks: GPT-5 showed a more elaborate reasoning ability, correctly calculating the number of floppy disks needed to install Windows 11, compared with GPT-4o, which relied more on general interpretations.
- Creative writing about Lincoln and basketball: GPT-5 developed a more coherent narrative, while GPT-4o limited itself to simpler, though still entertaining, answers.
- Kyle Orland's biography: GPT-5 produced a more detailed and up-to-date result, showing a greater web search ability than GPT-4o, which relied on static data.
- Emergency flight advice: GPT-4o provided more practical and clear guidance, while GPT-5 tried to condense the procedures, risking the omission of critical details.
Comparative analysis
Although GPT-5 won more rounds numerically, in many responses GPT-4o proved more detailed and personal, with more fully formed and in-depth answers. The preference between the models depends heavily on the use: GPT-5 tends toward quick and concise answers, while GPT-4o offers richer insights.
Conclusions
The comparison, although not exhaustive, highlights the challenges of developing a single model able to meet every need. GPT-5 represents a step forward in some areas but still has room for improvement, especially in consistency and reliability. The choice of model depends largely on the specific use and on individual preferences, with the risk of disappointing long-time users accustomed to earlier styles.
Ultimately, progress in artificial intelligence remains a balance between speed, accuracy and depth, and this comparison underlines how much remains to be done to reach a universal solution.