Recently, Wharton professor Ethan Morrick revealed the latest information about Anthropic. He said Anthropic's PR department has clarified that its flagship AI model, Claude3.7 Sonnet, costs as much as "ten millions of dollars" and uses less than 10^26 FLOP. Morrick also mentioned that Anthropic explicitly informed him that Sonnet3.7 will not be classified as the 10^26 FLOP model, but the model size will be significantly expanded in the future. Although TechCrunch has contacted Anthropic to confirm the information, no reply has been received as of press time.
Previously, Anthropic CEO Dario Amodei revealed that the training cost of Claude 3.5 (the predecessor of Sonnet) is also tens of millions of dollars when it was released in fall 2024. This cost seems quite favorable compared to the top model in 2023. For example, OpenAI spent more than $100 million on developing GPT-4, while Stanford University research estimates that Google's cost of training Gemini Ultra models is close to $200 million.
However, Amodei expects future AI models to cost billions of dollars, and this figure does not include additional work such as security testing and basic research. Furthermore, as the AI industry gradually adopts “inference” models that can solve problems for a long time, the computational cost of running these models may continue to rise.