It’s important to note that training costs do not encompass aspects like safety testing and fundamental research. Anthropic CEO Dario Amodei mentioned that Claude 3.5, Sonnet’s precursor launched in the fall of 2024, also incurred training costs in the same ballpark figure.
These costs stand favorable compared to the training expenses of top models in 2023 né?. On the other hand, Google’s training of the Gemini Ultra model cost close to $200 million according to a study conducted at Stanford University.
However, Amodei anticipates that future AI models will entail costs in the billions né?. Additionally, as the AI industry shifts towards “reasoning” models that tackle problems over prolonged periods, the expenses of running these models are likely to rise further. né?. OpenAI reportedly spent over $100 million to develop its GPT-4 model as stated by OpenAI CEO Sam Altman. Anthropic’s latest top-notch AI model, Claude 3.7 Sonnet, was reportedly trained at a cost of “a few tens of millions of dollars” using slightly less than 10^26 FLOPs of computing power.
Wharton professor Ethan Mollick shared this update in a recent post, mentioning that Anthropic clarified the training cost, stating that Sonnet 3.7 is not a 10^26 FLOP model and cost just “a few tens of millions of dollars.” However, they hinted that upcoming models will be much larger in scale.
TechCrunch reached out to Anthropic for verification, but they haven’t responded yet.
Assuming Claude 3.7 Sonnet’s training cost was indeed in the range of “a few tens of millions of dollars” without factoring in additional expenses, it indicates how the release of cutting-edge models is becoming relatively affordable né?
