Open AI, pioneer and leader in generative artificial intelligence (AI) since the launch of ChatGPT at the end of 2022, released a new model on Thursday as the race for ever more efficient AI assistants is in full swing.
“GPT-5 is like talking to a doctoral-level expert on any subject,” assured Sam Altman, boss of the American company, during a briefing for the press. He compared GPT-3, at work in the first version of ChatGPT, to a high school student who “sometimes gives the right answer, and sometimes answers anything,” and GPT-4 to a college student.
The technology giants are competing with new, ever more sophisticated models, capable of “reasoning” and performing tasks autonomously, with their sights set on so-called “general” or “superintelligent” AI, with cognitive capacities superior to those of humans.
The pace has picked up, as different groups seek to make their AI assistant – ChatGPT, Gemini (Google), Meta AI, Claude (Anthropic) etc. – essential on a daily basis for as many users and developers as possible.
Google and Meta (Facebook, Instagram) capitalize on their huge user bases. Anthropic has made a name for itself especially among professionals. Grok, from xAI, an Elon Musk company, is integrated directly into the social network But ChatGPT remains the name best known to the general public. It has nearly 700 million weekly active users.
“Super power”
OpenAI touted GPT-5 as its “smartest,” “fastest,” and “most useful” model to date. “GPT-5 can really do some awesome things for you. He can instantly create software, on demand (…) it’s an incredible superpower,” praised Sam Altman.
One of its engineers, Yann Dubois, asked the AI assistant in everyday language to create an online application to learn French with games. GPT-5 immediately produced hundreds of lines of code and the basic website was created in a few minutes.
According to the company, GPT-5 is also less prone to “hallucinations”: it “admits” when it doesn’t know instead of inventing a seemingly convincing answer.
Security has also been strengthened. “Before, the approach was binary. If the request seemed safe, the model executed, and otherwise it refused,” explained Alex Beutel, a product security manager at the company. With GPT-5, in case of doubt about potential criminal motivations, “the model will be limited to general information, which cannot cause harm”.
The assistant must also become customizable – the user will be able to choose a concise, caring or sarcastic tone – and will be able to connect to Gmail.
“This remains a language model, it is not a radical break,” developer Simon Willison commented on his blog Thursday, after two weeks of testing. “But he rarely makes a mistake and generally gives the impression of mastery, and sometimes even excellence.”
“Dizzying costs”
Microsoft, the main investor in OpenAI, announced that GPT-5 is now available on its various platforms for engineers. “OpenAI is going to eat Microsoft whole,” Elon Musk reacted on
OpenAI was founded as a nonprofit in 2015 – Elon Musk was a co-founder – with the mission of creating general AI that benefits all of humanity.
GPT-5 “is clearly an intelligent model, but it is missing something important to achieve general AI (…) it is not a model that continuously learns from the things it discovers,” said Sam Altman. “Obviously, we will have to invest in computing power at staggering costs to achieve this, but we intend to continue,” he added. The star Silicon Valley start-up was valued at 300 billion in March.
Microsoft, Google, Meta and Amazon each plan to spend tens of billions of $ on AI infrastructure (sophisticated data centers), cutting-edge chips and human resources this year.
