OpenAI Releases Most Powerful Model GPT-5, Comparable to a Doctor
OpenAI Releases GPT-5
Tech News, Beijing Time August 8th, OpenAI held a press conference today morning and officially released the highly anticipated new-generation large-scale language model GPT-5, which will be available to all 7 billion ChatGPT users.
OpenAI claims that GPT-5 is the company's most powerful AI system yet, with intelligence levels surpassing all previous models in multiple fields such as programming, mathematics, writing, healthcare, and visual perception.
OpenAI CEO Sam Altman said that GPT-5 and OpenAI's previous models have made significant progress. He compared it to the experience of using the first iPhone with a retina display, saying that it feels like you're talking to an expert or a Ph.D.
GPT-5
"GPT-5 really makes me feel like our main model has reached the level where you can ask it any question and get a response from a true expert or Ph.D.," Altman said at the press conference. "One of its coolest abilities is that it can write high-quality code for you on the spot. The concept of 'on-demand software' will become a hallmark feature of the GPT-5 era."
Unified System
GPT-5 is a unified system that presents itself as a single model, rather than being divided into a common model and a separate reasoning model like previous models.
It consists of three key components: a smart foundation model that can answer most questions; a deep reasoning model (GPT-5 Thinking) for solving more complex problems; and an intelligent router (smart routing system) that quickly determines which model to use based on conversation type, complexity, tool requirements, and user prompts.
This router will continuously learn and optimize through real-time feedback, including user behavior, response preferences, and accuracy metrics, to improve performance over time.
Most Powerful Programming Model
OpenAI claims that GPT-5 is the company's most powerful programming model yet. It excels in generating and debugging large-scale resources libraries. It can typically create visually appealing and responsive websites, applications, and games through a single prompt, thanks to its keen insight into aesthetics.
GPT-5 Programming Score
Early testers have also noted that it outperforms previous models in design decision-making, with a deeper understanding of spacing, font layout, and whitespace.
Altman said that GPT-5 is the "world's strongest programming and writing model"."
In OpenAI's testing, this model outperformed all other models in the programming test on the SWE-Bench, SWE-Lancer, and Aider Polyglot benchmarks. In real-world programming tests, GPT-5 scored 74.9% on SWE-bench Verified and 88% on Aider Polyglot.
At the press conference, Yann Dubois, responsible for post-training work, demonstrated how to use GPT-5 to generate a French learning website with interactive games.
Multimodal
OpenAI claims that GPT-5's multimodal capabilities have also been enhanced. This model excelled in multimodal benchmark tests, covering visual, video, spatial, and scientific reasoning fields.
Multimodal Test
Security Improvements
GPT-5 security research leader Alex Beutel said that OpenAI conducted over 5,000 hours of risk testing on GPT-5, focusing on ensuring that the model will not deceive users.
"In the past, we found that models would sometimes claim they had completed a task but actually hadn't. This is a problem."
"If someone asks 'What energy is needed to ignite a specific material?,' this might be an attempt to bypass safety protection mechanisms, trying to cause harm, or it could be a student asking a question for learning about physics. This presents a real challenge for the model on how to provide the best response."
OpenAI will release GPT-5 to all free users and paid ChatGPT subscribers starting Thursday, with educational and enterprise customers expected to gain access by next week. Paid users will enjoy higher usage limits.(Author/Rainbow)