The new flagship model AI of Anthropic, Claude 3.7 Sonnet, cost “a few tens of millions of dollars” to train using less than 10 ^ 26 flops of computing power.
It is according to Professor Wharton Ethan Mollick, who, in an X post, relayed a clarification he had received from the Professor of Anthropic on Monday. “I was contacted by Anthropic who told me that Sonnet 3.7 would not be considered a flop model of 10 ^ 26 and would cost a tens of millions of dollars”, ” He wrote“Although future models will be much larger.”
Techcrunch contacted Anthropic for confirmation, but had not received an answer at the time of publication.
Assuming that the Sonnet Claude 3.7 costs only “a few tens of millions of dollars” to train, and without taking into account related expenses, it is a sign of relatively cheap to publish advanced models. The predecessor of Claude 3.5 Sonnet, released in the fall of 2024, Cosser also a few tens of millions of dollars to trainThe CEO of Anthropic, Dario Amodei, revealed in a recent test.
These totals compare themselves quite favorably to the price labels for training of Top Models of 2023. To develop its GPT-4 model, Openai spent more than $ 100 million, according to to Openai CEO Sam Altman. Meanwhile, Google spent nearly $ 200 million to form its Gemini Ultra model, Stanford study estimated.
That said, Amodei expects future AI models Cost of billions of dollars. Admittedly, training costs do not capture work such as security tests and basic research. In addition, as the AI industry embraces the “reasoning” models that work on problems for long periods, the computer costs of the race models will probably continue to increase.
