Technology
Anthropic’s latest flagship AI model,Claude 3.7 Sonnet, is estimated to cost “around tens of hundreds of thousands of dollars” to train using less than 10^26 FLOPs of computational power.
This details comes from Wharton professor Ethan Mollick, who shared a clarification he received from Anthropic’s public relations in a post on X on Monday. “I was contacted by anthropic who informed me that Sonnet 3.7 woudl not be considered a 10^26 FLOP model and costs around tens of hundreds of thousands of dollars,” he stated, “even though future models are likely to be much larger.”
TechCrunch reached out to Anthropic for confirmation but had not received a reply by the time this article was published.
If Claude 3.7 Sonnet indeed costs approximately “tens of hundreds of thousands of dollars” for training alone, excluding related expenses, it indicates how relatively affordable it is becoming to launch state-of-the-art models. Claude 3.5, the predecessor to Sonnet released in fall 2024, also reportedly cost about tens of hundreds of thousands for training as noted by Anthropic CEO Dario Amodei in a recent article.
These figures compare favorably against the training costs associated with leading models from 2023. OpenAI invested over $100 million in developing its GPT-4 model according to OpenAI CEO Sam Altman,as reported. Meanwhile, Google allocated nearly $200 million for training its Gemini Ultra model as estimated by a Stanford study.
Nonetheless, Amodei anticipates that upcoming AI models willcost billions. Certainly, training expenses do not encompass tasks like security testing and essential research. Moreover, as the AI sector adopts “reasoning” models that tackle problems overextended durations, the computational costs associated with operating these models are expected to continue rising.
