The Claude 3.7 Sonnet, the latest flagship AI model for humanity, costs tens of millions of dollars to train using computing power less than 10^26.
It relayed the clarifications he received from humanity’s PR in a post on Monday, according to Professor Wharton Ethan Morrick. “We’ve been contacted by humanity who said that the Sonnet 3.7 is not considered a 10^26 flop model and would cost tens of millions of dollars,” he writes.
TechCrunch reached out to humanity to confirm, but had not received a response at the time of publication.
Assuming that the cost of Claude 3.7 Sonnet actually costs “tens of millions of dollars” to train without considering the related costs, how cheap it is to release cutting-edge models. It’s a sign. Sonnet’s predecessor, Claude 3.5, was released in the fall of 2024 and similarly cost tens of millions of dollars to train.
These totals are compared fairly favorably with the training price tag of the top model for 2023. According to Openai CEO Sam Altman, Openai spent more than $100 million to develop the GPT-4 model. Meanwhile, Google spent nearly $200 million to train Gemini Ultra models, Stanford’s research estimates.
That being said, Amodei expects future AI models to cost billions of dollars. Certainly, training costs do not acquire tasks like safety tests or basic research. Furthermore, the AI industry employs a “inference” model that addresses problems over a long period of time, which could lead to the computing costs of running models continuing to increase.
Source link
