Anthropic’s newest flagship AI model, Claude 3.7 Sonnet, cost “a few tens of millions of dollars” to train using less than 10^26 FLOPs of computing power.
That’s according to Wharton professor Ethan Mollick, who in an X post on Monday relayed a clarification he’d received from Anthropic’s PR. “I was contacted by Anthropic who told me that Sonnet 3.7 would not be considered a 10^26 FLOP model and cost a few tens of millions of dollars,” he wrote, “though future models will be much bigger.”
TechCrunch reached out to Anthropic for confirmation but hadn’t received a response as of publication time.
Assuming Claude 3.7 Sonnet indeed cost just “a few tens of millions of dollars” to train, not factoring in related expenses, it’s a sign of how relatively cheap it’s becoming to release state-of-the-art models. Claude 3.5, Sonnet’s predecessor, released in fall 2024, similarly cost a few tens of millions of dollars to train, Anthropic CEO Dario Amodei revealed in a recent essay.
Those totals compare pretty favorably to the training price tags of 2023’s top models. To develop its GPT-4 model, OpenAI spent more than $100 million, according to OpenAI CEO Sam Altman. Meanwhile, Google spent close to $200 million to train its Gemini Ultra model, a Stanford study estimated.
That being said, Amodei expects future AI models to cost billions of dollars. Certainly, training costs don’t capture work like safety testing and fundamental research. Moreover, as the AI industry embraces “reasoning” models that work on problems for extended periods of time, the computing costs of running models will likely continue to rise.
