The post on the OPENAI blog claims that the GPT-5 is overcoming its previous models on several coding reference studies, including in verified Benchie Ben (gaining 74.9 percent), Swo-Banner (GPT-5 thinking got 55 percent), and Aider Polyglot (obtained 88 percent), which test the model’s ability to set the model, complete tasks in style Freelance and operate in many multiple programs.
During the press briefing on Wednesday, the main after the OpenAI training, Yann Dubois, prompted GPT-5 to “create a beautiful, highly interactive application for my partner, an English speaker to learn French.” He ordered AI to include functions such as Daily Progress, various classes such as Flashcards and Quizy, and noticed that he wanted the application to be wrapped in a “very addictive motive”. After about a minute, an application generated by AI jumped out. Although it was only one demo on rails, the result was an elegant side, which provided exactly what Dubois asked for.
“He is a great coding collaborator and also stands out in agency tasks,” says Michelle Pokrass, leading after training. “It effectively performs long chains and tools [which means it better understands when and how to use functions like web browsers or external APIs]He warns the detailed instructions and presents in advance the explanations of his actions. “
Opeli also says in his blog post that GPT-5 is “our best model for health questions.” In three LLM references related to OpenAI-Hardbench, hardbench and health consensussystem card (A document describing the technical capabilities of the product and other research results) states that GPT-5 thinking exceeds previous models “with a significant margin”. The GPT-5 version thought was 25.5 percent at Healthbench, compared to 31.6 percent of O3. These results are approved by two or more doctors, according to the system card.
According to POKSSS, the model also has less hallucination, a common problem for artificial intelligence, in which it contains false information. Openai Alex Beutel security manager adds that “they have significantly reduced the fraud rate in the GPT-5”.
“We have taken steps to reduce the GPT-5-meter tendency to cheat, cheat or hack problems, although our soothing is not perfect and further tests are needed,” says the system card. “In particular, we trained the model so that he would seek gracefully when he was set with tasks that he could not solve.”
The company’s system card claims that after testing the GPT-5 models without access to viewing websites, scientists found their hallucinations indicator (which they defined as “the percentage of actual claims containing petite or main errors”) 26 percent less common than the GPT-4O model. GPT-5 thinking has a 65 % hallucination indicator compared to O3.
In the case of hints that can be double (potentially harmful or mild), Beuttel claims that the GPT-5 uses “safe and sound finishes”, which prompts the model to “give a possible response, but as part of restrictions safe and sound”. According to Beuttel, Opeli made over 5000 hours of red teams and tested with external organizations to make sure the system is solid.
Opeli claims that it now offers almost 700 million active CHATGPT users, 5 million paying business users and 4 million programmers using the API.
“The vibrations of this model are really good and I think that people will really feel it,” says ChatgPT head Nick Turley. “Especially the average people who did not spend time thinking about models.”