The arrival of GPT-5 has brought not only technical advancements but also heated debates within the user community. While the new model offers enhanced multimodal capabilities, safer operation, and greater customization, many users miss the more personal and engaging tone of its predecessors. The removal and swift return of GPT-4o has highlighted that emotional connection can be just as important as raw performance.

The release of GPT-5 was preceded by a long period of anticipation, during which its launch was repeatedly delayed. According to the announcement, OpenAI placed greater emphasis than ever before on testing the model’s safety and reliability. They also emphasized that AI alignment played a central role, aiming to ensure the model operates in a controlled manner consistent with the developers’ intentions. Together, these factors contributed to the company postponing the model’s deployment several times. This extended preparation period naturally created high expectations among both experts and everyday users.

One of the most significant innovations in GPT-5 is its departure from the “multi-model” approach that characterized previous generations. While GPT-4, GPT-4 Turbo and GPT-4o were each optimized for different tasks such as quick responses, deep reasoning or multimodality, GPT-5 aims to integrate these diverse capabilities into a single unified system. The idea is that users no longer need to manually choose between different models or modes; instead, GPT-5 automatically selects the most suitable specialized submodel for a given request behind the scenes. This allows the system to handle both everyday fast interactions and complex reasoning-intensive tasks seamlessly, presenting itself as a single “face” of intelligence to the user.

Behind this adaptive behavior are several innovations aimed at developers. GPT-5 introduces options (through API use) that allow the style and level of detail in its responses to be more precisely tailored to specific use cases. Parameters such as “verbosity” and “reasoning_effort” make it possible for a customer service chatbot to provide concise and direct answers, while an educational assistant can deliver more detailed and analytical responses. This flexibility not only streamlines the development process but also makes it easier to integrate the model smoothly into different workflows. In addition, coding performance has improved noticeably. On professional benchmarks, GPT-5 has significantly outperformed its predecessors, especially in complex tasks involving multiple programming languages.

According to its official communications, the GPT-5 development team placed strong emphasis not only on functionality but also on improving reliability. OpenAI highlights that one of the primary goals of the new generation was to reduce hallucinations, meaning the generation of false or fabricated information. This issue has long been a point of criticism not only for the company’s earlier models but also for nearly all competing systems. Particularly in sensitive areas such as medical advice, programming, or mathematical problem solving, GPT-5 has made significant improvements. According to early studies and benchmark tests, the rate of incorrect or inaccurate responses has dropped considerably, and in its detailed, reality-focused reasoning mode it makes up to one-sixth fewer errors compared to the previous generation.

This is partly the result of a new safety mechanism called “safe completions,” which allows the model not only to reject uncertain or potentially harmful prompts but also to restrict how it responds based on higher-level safety considerations. This can include providing a partial or harmless answer or clearly explaining why it cannot give a precise response. This approach has proven especially useful in dual-use domains, where information could be applied in both legitimate and harmful ways, and it measurably reduces the rate of errors, misleading statements, or excessive overconfidence in the model’s responses.

Speaking of tone, user experiences with the model so far have been highly mixed. While it is widely acknowledged that the GPT-5 represents a significant technical leap forward, community feedback suggests a different story when it comes to its personality. Many users feel that the model is less creative, more distant, and even somewhat sterile compared to its predecessor, GPT-4o. Across platforms such as Reddit and X, people have expressed nostalgia for the friendlier, warmer tone of earlier versions, noting that GPT-5’s replies often feel shorter, less personal, and less engaging. Some have gone so far as to label it a “beige zombie” or even a step backward. Criticism has also grown over GPT-5’s handling of longer contexts, with many noting that it is less consistent, more prone to asking clarifying questions, or avoiding taking a firm stance. These reactions highlight that technical progress alone is not always enough, and that human factors such as personality, warmth, and a natural tone are equally important for user satisfaction.

A major source of frustration was that GPT-4o, valued for its lighter, more creative, and approachable style, was removed from the options available to ChatGPT Plus subscribers after the update. For many who had grown attached to that tone, its disappearance felt not just like a technical change but a personal loss.

Emotional reactions spread quickly across social media and online forums. OpenAI CEO Sam Altman later admitted that the team had underestimated the importance of stylistic differences and noted that for many users these systems are not merely tools but also sources of social connection. The backlash eventually grew to the point where OpenAI quickly announced that GPT-4o would once again be available alongside GPT-5.

The events highlighted that the success of an AI model is not determined solely by its technical specifications. Just as important is preserving familiar interaction styles, maintaining user trust, and fostering emotional attachment. While GPT-5 represents progress in several areas, for many it felt more like gradual evolution than a revolutionary breakthrough.

The short history of GPT-5 illustrates that artificial intelligence development today is no longer merely a technological race. Although the model brought tangible improvements in areas such as multimodal capabilities, flexible parameter tuning, reliability, and safety, user experience, familiar style, and emotional connection proved to be equally decisive factors. The removal and swift reinstatement of GPT-4o, along with the open community debate, make it clear that evaluating an AI system cannot be confined to performance metrics alone. People often judge technology through the lens of their relationship with it, their emotions, and their habits.

The capabilities and reception of GPT-5 symbolize both a more mature and safety-oriented development path and the recognition that AI also has a “human side.” In the future, the most successful systems are likely to be those that combine engineering sophistication with an understanding of users’ emotional and social needs. For OpenAI, GPT-5 is not only a milestone but also a lesson: the next step must involve not only a technological leap but also an elevation of user trust and personal experience to a new level.

István ÜVEGES, PhD is a Computational Linguist researcher and developer at MONTANA Knowledge Management Ltd. and a researcher at the HUN-REN Centre for Social Sciences. His main interests include the social impacts of Artificial Intelligence (Machine Learning), the nature of Legal Language (legalese), the Plain Language Movement, and sentiment- and emotion