Free unlock edited abstracts
FT’s editor Roula Khalaf chose her favorite stories in this weekly newsletter.
Openai has launched its largest AI model with fewer “illusions”, the latest version of a series of AI releases by the American Tech Group.
The San Francisco-based company unveiled its GPT-4.5 on Thursday, with a long-standing update to its technology update for its popular product, Chatgpt. In early testing, its hallucination rate (the AI system produces inaccurate information) was 37%, while its predecessor, GPT-4O, had an hallucination rate of nearly 60%.
With GPT-4.5, OpenAI continues to bet on large, expensive large language models, despite the emergence of powerful smaller products, such as China’s launch of DeepSeek’s R1, which are open, cheaper and easier to access for developers.
As competition in the rapidly growing AI industry is growing, technology groups are rushing to launch their latest models in recent weeks. Anthropic revealed its Claude 3.7 sonnet on Monday, the Grok 3 launched last week, the latest model from the Elon Musk Xai.
Openai said in a blog post Thursday that GPT-4.5 “has a wider knowledge and deeper understanding of the world, resulting in reduced hallucinations and more reliable in a wide range of topics.”
“With every new order of calculations, novel features will emerge,” the company said, adding that GPT-4.5 is the “possible boundary in unsupervised learning.”
Openai has been in a global competition to lead the AI industry, raising billions of dollars from investors to fund larger models, with increased functionality and requiring a lot of computing power.
It is in talks with SoftBank and other investors to raise up to $40 billion in valuation, including new funds. Two people who understand the process said Anthropic also raised about $3.5 billion at a $6 billion valuation.
However, the huge cost of running larger numbers has led Openai to consider withdrawing access to GPT-4.5 amid fierce competition for competitor pricing.
Openai said that while a preview of GPT-4.5 will be provided to developers using the OpenAI model through its application programming interface (API), this access can be revoked in the future.
AI groups generate revenue in large part through paid API access and personal subscriptions. Openai said it would consider how developers can use a powerful model and whether it is worth providing them with considering the high cost of running.
The company said: “GPT-4.5 is a very large compute-intensive model that makes it more expensive than the GPT-4O rather than a replacement for the GPT-4O [its predecessor]. Therefore, we are evaluating whether we want to continue serving it in the API for a long-term basis, as we are balanced with supporting current functionality with building future models. ”
OpenAI CEO Sam Altman has previously stated that the training cost of GPT-4 exceeds $100 million, and this cost is generally increased with the size and functionality of the model and requires more computing power to train and run.
In an article on X after the announcement, Altman said the company is “not in the GPU” to run and train the chips needed for AI systems.
“This is not the way we want to operate, but it’s hard to fully predict the growth wave that leads to GPU shortages,” he said, adding that they hope to get more in the coming weeks.
He also stressed that GPT-4.5 is not focused on reasoning and will not beat industry benchmarks, but it has “another wisdom that I’ve never felt magic before.”
George Hammond’s other reports in San Francisco and Melissa Heikkilä in London
