Unlock the publisher’s digest free
Roula Khalaf, editor -in -chief of the FT, selects her favorite stories in this weekly newsletter.
Openai launched its greatest model of artificial intelligence with fewer “hallucinations”, the latest of a wave of AI versions by American technological groups.
Thursday, the company based in San Francisco unveiled GPT-4.5, its long-awaited update of the technology underlying its popular Chatgpt product. In the first tests, its hallucination rate, where AI systems generate inaccurate information, was 37% against almost 60% on its GPT-4O predecessor.
With GPT-4.5, Openai continues to bet on large costly large models despite the advent of smaller products very competent, such as the Chinese start-up Deepseek’s R1, which are open, cheaper and more accessible for developers.
This occurs while competition is developing in the AI industry in rapid development, technological groups, while technology groups have rushed to launch their latest models in recent weeks. Anthropic revealed his Sonnet Claude 3.7 on Monday, which followed the launch of Grok 3 last week, the last model of the XAI of Elon Musk.
OpenAi in a blog article said Thursday that GPT-4.5 had “wider knowledge and a more in-depth understanding of the world, resulting in a reduction in hallucinations and more reliability in a wide range of subjects”.
“With each new order of size of calculation comes from new capacities,” said the company, adding that GPT-4.5 was “at the border of what is possible in unopensed learning”.
OPENAI was at the forefront of a world race to direct the AI industry, increasing dozens of billions of investors to finance larger models with increased capacities that require large amounts of computer power.
He is in talks with SoftBank and other investors to collect up to $ 40 billion to an assessment of $ 300 billion, including new money. Anthropic also collects around $ 3.5 billion in an evaluation of $ 60 billion, two people who know this process said.
However, the enormous costs of the management of larger models led OpenAI to consider withdrawing the access of developers to GPT-4.5 in the middle of fierce competition on competitors’ prices.
OPENAI said that if an overview of GPT-4.5 will be made available to developers who pay to use OPENAI models via its application programming interface (API), this access could be revoked in the future.
The AI groups largely generate revenues thanks to access to the paying API and individual subscriptions. Openai said he will see how the developers use the powerful model and if it was worth offered by considering the high cost of execution.
The company said: “GPT – 4.5 is a very large and high calculation intensity model, which makes it more expensive than and not a replacement of GPT – 4O [its predecessor]. For this reason, we assess the opportunity to continue to serve it in the long term of the API, while we will balance the current capacity for supporting the construction of future models. »»
Openai CEO Sam Altman said GPT-4 cost more than $ 100 million to train, and these costs should increase as the size and capacity of the models evolve and require more computing power to train and operate.
In an article on X after the announcement, Altman said that the company was “outside GPU”, the chips necessary to execute and train AI systems.
“This is not how we want to operate, but it is difficult to predict growth overvoltages that lead to GPU shortages,” he said, adding that they expect to get more in the coming weeks.
He also pointed out that GPT-4.5 was not focused on reasoning and would not beat the landmarks of the industry, but there was “another type of intelligence and there is a magic that I have never felt before”.
Additional report by George Hammond in San Francisco and Melissa Heikkilä in London
