Over a year has passed since OpenAI released ChatGPT 4, and although GPT 4 Turbo was eventually released, it was more of an improvement over the original GPT 4. For this reason, a lot of people anticipate seeing a legitimate successor this year, although it won’t be GPT 5.
Analysts predict that GPT 5 won’t occur this year due to the enormous amount of processing power needed for such a model.
According to Dan Hendrycks, the director of the Center for AI Safety, OpenAI need 10 times more processing power to improve its GPT LLM. Therefore, if GPT-5 were to replace GPT-4.5, about a hundred times more processing power would be required than for GPT-4. That is equivalent to operating around one million H100 chips nonstop for three months, to put things into perspective. The analyst questions whether OpenAI already has a server with one million GPUs.
The information provided by Dario Amodei, CEO of Anthropic, supports this assertion. He recently brought attention to the fact that educating a state-of-the-art LLM now costs about $1 billion. But by 2025 or 2026, this cost is expected to increase to $5 billion to $10 billion. It is noteworthy that the $1 billion training cost, which is a plausible assumption for GPT-4.5, reflects the tenfold increase in computational resources.
Given that GPT 4.5 has already surfaced in leaks from the OpenAI blog, its release this year is more plausible. A site from the OpenAI blog describing GPT 4.5, which is anticipated to be quicker, more accurate, and more scalable, appeared in search engine results last month.
