Unlock the editor of the editor for free
FT’s editor, Roula Khalaf, chooses your favorite story this week.
Openai has launched the largest AI model with the latest “hallucinations” of the US Technology Group’s AI release.
Headquartered in San Francisco, the company unveiled GPT-4.5 on Thursday. GPT-4.5 has been updated for a long time for the technology that supports the popular product CHATGPT. In the initial test, the hallucinations of the AI system generated inaccurate information were 37 %compared to almost 60 %of the previous model GPT-4O.
With GPT-4.5, Openai continues to bet on large and expensive large language models despite the emergence of small-scale products with the same ability as the R1 of China Start Up DeepSeek.
As the technology group rushes to launch the latest models in recent weeks, competition is growing in the rapidly developing AI industry. Anthropic unveiled Claude 3.7 Sonnet on Monday and launched Grok 3, the latest model of ELON Musk’s XAI last week.
In a Thursday blog post, OpenAI said GPT-4.5 said, “It has a wide range of knowledge and a deeper understanding of the world, causing hallucinations and more reliability of various topics.”
The company added GPT-4.5.
Openai stood at the forefront of the world race to lead the AI industry and raised tens of billions of dollars from investors to fund the larger models to enhance the ability to require vast amounts of computing power.
Taking up to $ 400 billion in a $ 3 billion evaluation, including soft banks and other investors, we talk with Softbank and other investors. Anthropic is also raising a $ 3.5 billion fund for more than $ 6 billion, said the two with knowledge of the process.
However, due to the high cost of operating a larger model, Openai considered withdrawing access to GPT-4.5 in the fierce competition of rival prices.
Openai said the developer pays for the use of Openai’s models through the Application Programming Interface (API) that it will provide a preview of GPT-4.5.
The AI group generates a lot of money through paid API access and individual subscriptions. Openai said that it will see how developers use powerful models and whether they are worth providing in consideration of high costs.
The company said: “GPT-4.5 is more expensive than replacing GPT-4Os with a very large and computing-intensive model. [its predecessor]. As a result, we are evaluating whether we will continue to service in the long term because we will build the future model and balance the current function. ”
Openai CEO SAM ALTMAN costs more than $ 100 million in GPT-4 training, and this cost is expected to increase widely depending on the size and function of the model size and require more computing power for training and operation.
In a post about X after the announcement, Altman said that the company is out of GPU, a chip that the company needs to operate and train the AI system.
“This is not the way we want to run, but it is difficult to predict the growth surge in the GPU shortage.”
He also emphasized that GPT-4.5 did not focus on reasoning and failed to overcome the industry benchmark, but “I had different kinds of intelligence and there was magic that I had never felt before.”
Additional report from George Hammond in San Francisco and Melissa Heikkilä in London