Anthropic's New AI: Surprisingly Affordable to Train

# Claude

Anthropic's latest AI model, Claude 3.7 Sonnet, was reportedly trained for just a few tens of millions of dollars, using less than 10^26 FLOPs of computing power. This info comes from Wharton professor Ethan Mollick, who shared on X that he got this straight from Anthropic's PR team. They told him, "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," though they hinted that future models would be way bigger. TechCrunch tried to get the scoop from Anthropic but didn't hear back before their article went live. If Claude 3.7 Sonnet really did cost just a few tens of millions to train, it's pretty wild to think how much cheaper it's getting to roll out top-notch AI models. Just to give you an idea, Claude 3.5, which came out in the fall of 2024, also cost a few tens of millions to train, according to Anthropic's CEO, Dario Amodei. When you stack these numbers against what it cost to train the big dogs from 2023, it's even more impressive. OpenAI dropped over $100 million to get their GPT-4 model up and running, according to their CEO, Sam Altman. And Google? They spent close to $200 million on their Gemini Ultra model, if you believe a Stanford study. But hold on, Amodei thinks the price tag for future AI models could hit the billions. And remember, training costs are just one piece of the puzzle. There's also safety testing, fundamental research, and all that jazz. Plus, as the AI world starts leaning into "reasoning" models that chew on problems for longer, the costs to keep these models running are probably gonna keep climbing.

Related article

Comments (32)