Anthropic’s latest flagship AI mannequin, Claude 3.7 Sonnet, value “just a few tens of tens of millions of {dollars}” to coach utilizing lower than 10^26 FLOPs of computing energy.
That’s in line with Wharton professor Ethan Mollick, who in an X publish on Monday relayed a clarification he’d obtained from Anthropic’s PR. “I used to be contacted by Anthropic who advised me that Sonnet 3.7 wouldn’t be thought of a ten^26 FLOP mannequin and price just a few tens of tens of millions of {dollars},” he wrote, “although future fashions can be a lot greater.”
TechCrunch reached out to Anthropic for affirmation however hadn’t obtained a response as of publication time.
Assuming Claude 3.7 Sonnet certainly value simply “just a few tens of tens of millions of {dollars}” to coach, not factoring in associated bills, it’s an indication of how comparatively low-cost it’s changing into to launch state-of-the-art fashions. Claude 3.5 Sonnet’s predecessor, launched in fall 2024, equally value just a few tens of tens of millions of {dollars} to coach, Anthropic CEO Dario Amodei revealed in a current essay.
These totals examine fairly favorably to the coaching worth tags of 2023’s high fashions. To develop its GPT-4 mannequin, OpenAI spent greater than $100 million, in accordance to OpenAI CEO Sam Altman. In the meantime, Google spent near $200 million to coach its Gemini Extremely mannequin, a Stanford examine estimated.
That being mentioned, Amodei expects future AI fashions to value billions of {dollars}. Definitely, coaching prices don’t seize work like security testing and elementary analysis. Furthermore, because the AI business embraces “reasoning” fashions that work on issues for prolonged intervals of time, the computing prices of working fashions will doubtless proceed to rise.
