Unlock the digestive of free editor
Roula Khalaf, the FT editor, chooses her favorite stories in this weekly newsletter.
Openai has begun his largest model of artificial intelligence with fewer “hallucinations”, the latest in a variety of omissions of it from American technology groups.
The San Francisco-based company on Thursday discovered the GPT-4.5, its long-awaited updating for the technology that supports its well-known chatgpt product. In the early tests, its hallucination rate, where systems it generates incorrect information, was 37 percent compared to nearly 60 percent in its GPT-4o ancestors.
With GPT-4.5, Openai is still betting on large, expensive models of large languages, despite the arrival of smaller very capable products, such as the Chinese start-to-day R1, which are open, cheaper and more accessible to developers.
Coming as competition is growing within the developing industry as technology groups have rushed to start their latest models in recent weeks. Anthropic discovered its Sonet Claude 3.7 on Monday, which followed the launch of Grok 3 last week, the latest model by Elon Musk Xai.
Openai in a blog post on Thursday said that the GPT-4.5 had “wider knowledge and a deeper understanding of the world, leading to decreased hallucinations and more reliability in a wide range of topics”.
“With any new order of the size of the calculation comes new skills,” the company said, adding GPT-4.5 was “to the limit of what is possible in unseen teaching.”
Openai has been in the vanguard of a global competition to lead him industry, raising tens of billions of investors to finance larger models with adult skills that require large amounts of computing power.
It is in talks with softbank and other investors to raise up to $ 40 billion in a $ 300 billion rating, including new money. Anthropic is also collecting funds about $ 3.5bn in a $ 60bn-plus rating, two people with knowledge of that process said.
However, the high costs of running the larger models have led Openai to consider the attraction of developer entry into the GPT-4.5 in the midst of fierce competition for prices from rivals.
Openai said while a GPT-4.5 survey will be made available to developers paying to use OpenAI models through its application programming interface (API), this approach can be revoked in the future.
Groups of it produce mainly income through paid access to API and individual reconciliation. Openai said he will see how developers use the powerful model and if it is worth offering them taking into account the high cost of his direction.
The company said: “GPT-4.5 is a very large and calculator model, making it more expensive than and not a replacement for GPT-4o (its predecessor). Because of this, we are assessing whether we will continue to serve the long-term api as we balance the support of current skills with the construction of future models.”
Openai chief Openai Sam Altman has previously said the GPT-4 costs more than $ 100 million to train, and such costs are expected to increase widely as the size and skills of models on the scale and require more computing power to train and run.
In a post on X after the announcement, Altman said the company was “out of GPU”, chips required to run and train him systems.
“This is not how we want to operate, but it’s hard to predict the perfect growth growing that leads to GPU deficiency,” he said, adding that they were expecting to get more in the coming weeks.
He also emphasized the GPT-4.5 was not focused on reasoning and would not overcome industry standards, but it had “another type of intelligence and there is a magic for it I have not felt before.”
Additional reporting by George Hamond to San Francisco and Melissa Heikilä in London
