The last flagship AI of Anthropic may not have been incredibly expensive to train

The most leading AI model of Anthropic, Claude 3.7 Sonnet, cost “several tens of millions of dollars” to train using less than 10^26 flops of computing power.

This is, according to Professor of Wharton Ethan Molek, who in the publication X on Monday prevails that he received from the PR of Anthropic. “I contacted me with Anthropic, who told me that Sonnet 3.7 would not be considered a 10^26 flop model and would cost several tens of millions of dollars,” ” He wrote“Although future models will be much more.”

TechCrunch turned to anthrop for confirmation, but did not receive an answer to the publication time.

Assuming that Claude 3.7 Sonnet really cost only a few tens of millions of dollars to train, not to be reported in related expenses, this is a sign of how relatively cheap it is to put the most up-to-date models. Claude 3.5, Sonnet’s predecessor issued in the fall of 2024, Similarly costs several tens of millions of dollars for trainingThe anthropic CEO Dario Amodei revealed in a recent essay.

These amounts compare quite favorably with the prices of training for the best models of 2023. To develop its GPT -4 model, Openai spends more than $ 100 million, According to For Openai CEO Sam Altman. Google, meanwhile, spends nearly $ 200 million to train its model Gemini Ultra, Stanford Study calculatedS

Given AModei expects future AI models to costs billions of dollarsS Certainly, training costs do not capture work as a testing of safety and fundamental research. Moreover, as the AI ​​industry covers the “reasoning” models that work on problems with Extended periods of timeComputer costs for working models will probably continue to grow.