OpenAI has officially announced the release of GPT-5, the latest flagship model, which will be accessible to all users of ChatGPT, including developers.
According to CEO Sam Altman, GPT-5 represents a significant advancement over earlier versions. He likened its transformative nature to that of the first iPhone equipped with a Retina display, remarking, “It’s something I just don’t want to go back from.”
OpenAI touts GPT-5 as being smarter, faster, and more accurate. During a recent press briefing, Altman described the evolution of their models: “GPT-3 felt like conversing with a high school student. With GPT-4, it was more like a college student. Now, GPT-5 truly resembles a PhD-level expert.”
Despite ChatGPT having nearly 700 million weekly users, OpenAI has not released an industry-leading model in some time. The launch of GPT-5 is expected to regain its competitive edge in the market. Altman remarked, “This is the leading model worldwide for coding, writing, healthcare, and numerous other fields.”
One notable change with GPT-5 is its presentation within ChatGPT. Users will see it as a single model rather than having separate reasoning functionalities. The model incorporates a router developed by OpenAI that seamlessly switches to a reasoning version when faced with more complex queries, or upon user prompting with “think hard.” Altman had previously criticized the model selection interface as a “very confusing mess.”
Nick Turley, head of ChatGPT, expressed optimism about the new model, stating, “The vibes of this model are really good. I believe users, especially those not deeply immersed in model discussions, will notice a positive change.”
OpenAI is making GPT-5 immediately available to all ChatGPT users; however, free users will encounter an undisclosed limit on prompts, beyond which the model will revert to a less powerful “mini” version. For developers using OpenAI’s API, GPT-5 will be offered in three variants — GPT-5, GPT-5 mini, and GPT-5 nano, each at different price points.
Additionally, OpenAI will introduce four new personality themes for ChatGPT, allowing users to customize responses. The themes include “Cynic,” “Robot,” “Listener,” and “Nerd,” along with options to modify the color of individual chat threads.
Altman anticipates that GPT-5’s advanced coding abilities will bring about an era he describes as “software on demand.” In testing conducted by OpenAI, the model outperformed others in coding across several benchmarks including SWE-Bench, SWE-Lancer, and Aider Polyglot.
During the press conference, Yann Dubois, OpenAI’s lead for post-training, demonstrated GPT-5’s capabilities by generating a study website for learning French that featured an interactive game. Within seconds, the model composed hundreds of lines of code, showcasing a fully functional frontend that Dubois displayed on Zoom.
Over five thousand hours of testing were conducted to evaluate GPT-5’s safety risks, according to Alex Beutel, who leads safety research for the model. A primary concern was ensuring the model does not provide false information. GPT-5 displays fewer hallucinations compared to OpenAI’s previous o3 reasoning model, though the issue of confidently providing inaccurate information persists for large language models.
Handling requests wherein the model operates like an agent can exacerbate this issue. However, OpenAI has indicated that GPT-5 is more adept at managing multi-step tasks. Beutel noted, “In the past, the model would claim it could complete a task it hadn’t actually accomplished, and this is problematic.”
GPT-5 is designed to provide “safe completions” for prompts that it would have previously rejected. For example, if requested to give the energy needed to ignite a specific material, the model may recognize that such knowledge could be misused but also sees the potential educational value of these inquiries. Beutel described this as a challenging area for the model.
With safe completions, GPT-5 aims to provide helpful answers while remaining within safety constraints. Beutel clarified, “The model will only partially comply, often focusing on high-level information that cannot be misused.”
Moreover, OpenAI asserts that GPT-5 now better recognizes when it is unable to fulfill a task or deliver an accurate response, a feature aimed at enhancing user trust. The specific training data used for GPT-5 remains undisclosed by the company.
“This is clearly a model that is generally intelligent”
OpenAI aims to develop artificial general intelligence (AGI), and Altman believes that GPT-5 moves closer to that aspiration, despite industry shifts toward “superintelligence.”
“I somewhat dislike the term AGI since it varies in meaning across the industry,” Altman remarked. “However, this model represents a significant advancement toward truly capable models. This is evidently a model with general intelligence.”
Nonetheless, he acknowledged that GPT-5 still lacks a crucial component.
“This model does not learn continuously from new experiences or data it encounters, which I believe is essential for AGI.”
