Training Sonnet’s latest flagship AI model, Claude 3.7, cost its developer, Anthropic, just “tens of millions of dollars” and required less than 1014 Tflops of computing power.
Image source: anthropic.com
This was reported by Ethan Mollick, a professor at the Wharton School of Business (USA), who cited an explanation given by Anthropic’s public relations department. “Anthropic representatives contacted me and said that Sonnet 3.7 should not be considered a 1026 Flops model, and it only cost a few tens of millions of dollars,” the scientist said, adding that “future models will be much larger.”
If we assume that training Anthropic’s Claude 3.7 Sonnet really did cost the company “only tens of millions of dollars” without the associated costs, then the cost of developing the systems is really starting to come down. Its predecessor, the mid-sized Claude 3.5 Sonnet, released last year, cost a similar amount to train, Anthropic CEO Dario Amodei said. By comparison, OpenAI spent $100 million to develop GPT-4, and Google spent an estimated $200 million to train Gemini Ultra.
Mr. Amodei, however, does not expect a long-term reduction in the cost of training AI — they will already cost billions of dollars, and that’s not counting the costs of security testing and fundamental research. “Reasoning” models are now coming into circulation, which means that AI will require more and more computing resources.