Switch off the editor’s digest free of charge
Roula Khalaf, editor of the FT, selects her favorite stories in this weekly newsletter.
Openaai has launched its largest model for artificial intelligence with fewer “hallucinations”, the latest in a flood of AI publications by US tech groups.
The company based in San Francisco presented GPT-4.5 on Thursday, its long-awaited update of technology, which underpins its popular product chatt. In early tests its hallucination rate, where AI Systems generate inaccurate information and amounted to 37 percent compared to almost 60 percent with its predecessor GPT-4O.
With GPT-4.5, Openai Still bet on large, expensive major language models, although highly delicious smaller products such as the Chinese start-up deepseeks R1, which are open, cheaper and more accessible to developers.
It happens that the competition in the rapidly developing AI industry is growing, as tech groups will start their latest models in recent weeks. Anthropic revealed his Claude 3.7 sonnet on Monday, which was followed last week by Grok 3, the latest model by Elon Musk’s Xai.
Openaai said on Thursday in a blog post said GPT-4.5, “a broader knowledge and a deeper understanding of the world, which led to reduced hallucinations and more reliability in a variety of topics”.
“With every new order of the calculation, new functions come,” said the company and added that GPT-4.5 “was possible at the border of what is possible in unattended learning”.
Openai was at the top of a global race to lead the AI industry and increased tens of billions of investors to finance larger models with increased skills for which large amounts of computing power are required.
It is in discussions with soft bank and other investors to collect up to $ 40 billion in an evaluation of $ 300 billionincluding the new money. Anthropic also collects a foundation of around $ 3.5 billion with $ 60 billion, said two people with knowledge of this process.
However, the enormous costs for the management of larger models have led Openai to be due to the developer access to GPT-4.5 under violent competition for the pricing of competitors.
Openaai said that developers who pay for the use of Openai models via the application programming interface (API) for the use of OpenAI models are provided with a preview of GPT-4.5. This access could be revoked in the future.
AI groups mainly generate income on paid API access and individual subscriptions. Openaai said it will see how developers use the powerful model and whether it is worth taking into account the high costs for the operation.
The company said: “GPT-4.5 is a very large and arithmetic model, which makes it more expensive than and not a replacement for GPT-4O (its predecessor). For this reason, we evaluate whether we should continue to serve you in the API in the long term, as we reconcile the support of the current skills in building future models. “
Sam Altman, Chief Executive from Openaai, has already said that GPT-4 cost more than $ 100 million for training, and these costs are generally expected if the size and skills of models scale and require more computing power for training and run.
In a post on X after the announcement, Altman said that the company was “out of the GPUS”, the chips that were necessary for the running and training of AI systems.
“We don’t want to work like this, but it is difficult to perfectly predict growth fluctuations that lead to GPU bottlenecks,” he said, adding that they would get more in the coming weeks.
He also emphasized that GPT-4.5 would not focus on the reasoning and that the industry benchmarks would not exceed, but it had “a different kind of intelligence and there is a magic that I didn’t feel before.
Additional reporting by George Hamond in San Francisco and Melissa Heikilä in London
