OpenAI Unveils GPT-5: A ‘PhD-Level Expert’ for Every User
OpenAI has unveiled GPT-5, its new flagship artificial intelligence model, now available to all ChatGPT users and developers. Company CEO Sam Altman described it as “a leap you won’t want to come back from,” likening its debut to the arrival of the first iPhone with a Retina display. According to him, GPT-5 is smarter, faster, and makes fewer mistakes than its predecessors. “GPT-3 was like talking to a high school student—sometimes correct, sometimes completely off. GPT-4 felt like a university student. GPT-5, for the first time, feels like a conversation with a PhD-level expert,” he said.
Despite ChatGPT’s nearly 700 million weekly users, OpenAI has not held a clear lead in the AI model race in recent years. The company believes GPT-5 will restore that leadership. “This is the best AI in the world for programming, writing, medicine, and much more,” Altman asserted.
In ChatGPT, the new engine is presented as a single unified model without a separate “reasoning mode.” Under the hood, a specialized router automatically activates an advanced mode for complex queries or when prompted with “think deeper.” Altman admitted that the previous system of choosing between models was “confusing and inconvenient.” ChatGPT head Nick Turley added: “This model has a great energy. People will feel it—especially those who don’t think about how AI models work at all.”
GPT-5 is available immediately, though free-tier users face a hidden cap on the number of requests; after exceeding it, the system switches to a lighter version. Through the API, developers can choose from three variants: GPT-5, GPT-5 mini, and GPT-5 nano. ChatGPT also introduces four new “personalities”—Cynic, Robot, Listener, and Nerd—as well as the option to change the color of individual conversations.
During the presentation, Altman called GPT-5 “the engine of the on-demand software era.” Company tests show the model achieving record results in several industry coding benchmarks. At the briefing, Yann Dubois, head of post-training, demonstrated GPT-5 by building an educational website for learning French with an interactive game—the AI generated hundreds of lines of code and immediately assembled a working interface.
Alex Boitel, head of safety research, noted that GPT-5 underwent more than 5,000 hours of testing, with particular attention to reducing false outputs. The model hallucinates less often than its predecessors, though the issue of confidently delivering inaccurate information has not been entirely eliminated. A new “safe answers” system allows the AI to provide maximally useful yet safe responses to potentially risky queries—for example, distinguishing between legitimate scientific questions and attempts to bypass safety measures.
GPT-5 is also better at recognizing when it cannot complete a task or provide an accurate answer, a change aimed at increasing user trust. OpenAI has not disclosed the data on which the model was trained. Altman believes GPT-5 brings the company closer to its goal of creating artificial general intelligence, though there is still ground to cover—the model still cannot learn from new data in real time.