OpenAI, the company at the forefront of generative artificial intelligence since the launch of ChatGPT in late 2022, unveiled its latest AI model on Thursday, labelling it the “most intelligent,” “fastest,” and “most useful” creation yet.
There was anticipation for the new model from the US company, with tech giants locked in a race to develop the most advanced AI assistants. The goal is artificial general intelligence (AGI) or artificial superintelligence (ASI), with cognitive abilities surpassing human capabilities.
Speaking at a press briefing, Sam Altman, Open-AI co-founder and CEO, compared GPT-5 to conversing with a PhD-level expert on any topic. On the other hand, GPT-3, used in the first version of ChatGPT, was like a school student who sometimes gets the answer right, and sometimes doesn’t, he said. He likened GPT-4 to a university student.
Companies leading in generative AI are competing with models that increasingly excel at “reasoning” and performing tasks autonomously. This is accelerating as various tech companies aim to make their AI assistant – ChatGPT, Gemini (Google), Meta AI, Claude (Anthropic) – essential for as wide an audience as possible.
ChatGPT remains the most recognised name among the public, boasting nearly 700 million active users per week. OpenAI aims to maintain this lead.
The AI assistant will soon be customisable, allowing users to choose between a concise, friendly, or sarcastic tone. It will also integrate with Gmail.
GPT-5 can do truly impressive things such as creating software immediately upon request, according to Altman.
One of the engineers, Yann Dubois, asked the AI assistant in plain language to develop an online app for learning French through games. GPT-5 subsequently produced hundreds of lines of code, creating a basic website in mere minutes.
According to OpenAI, GPT-5 is also more reliable, “hallucinating” less than earlier models and “admitting” when it does not know an answer instead of generating a convincing fabrication. “We have trained it to be honest, so it doesn’t lie to users,” said Alex Beutel, responsible for the safety of the company’s products.
It is also more difficult to use GPT-5 for dubious purposes. “Previously, there was a binary approach. If a request seemed safe, the model carried it out. Otherwise, it refused,” said the researcher. Now, if there is uncertainty regarding potential criminal intent, GPT-5 will “limit itself to general information that cannot cause harm.”
OpenAI was founded in 2015 as a non-profit organisation aiming to develop artificial general intelligence (AGI) for the benefit of all humanity. GPT-5 is clearly an intelligent model, Altman said. However, he acknowledged that it lacks something "crucial" for AGI: it is not a model that continually learns from the things it discovers.
“Certainly, staggering amounts will need to be invested in computing power to achieve that. But we intend to stay on that path,” he added.
