Archive — US — techcrunch
Anthropic’s latest flagship AI might not have been incredibly costly to train
Anthropic's Claude 3.7 Sonnet AI model was trained for 'a few tens of millions of dollars' using less than 10^26 FLOPs, indicating a decline in training costs for state-of-the-art models. Unlike 2023's costly models like OpenAI's GPT-4 and Google's Gemini Ultra, future models are anticipated to be significantly more expensive due to increased complexity.
Key points
- Claude 3.7 Sonnet's training cost is estimated at a few tens of millions of dollars.
- The FLOPs used for training are reported to be under 10^26, suggesting reduced computational demands.
- Previous models, such as Claude 3.5, had similar costs, showcasing a downward trend in training expenses.
- OpenAI and Google spent over $100 million and nearly $200 million, respectively, on their leading models.
- Anthropic's CEO indicates that while current costs are lower, future models might escalate to billions.
- Training costs exclude related expenses like safety testing and fundamental research.
Why it matters
- The trend of decreasing training costs aligns with the growing accessibility of computing power and advancements in AI efficiency.
- Historically, the AI industry has seen fluctuating training expenses, with significant investments in 2023 for leading-edge models.
- The shift to more complex reasoning models could drive up costs despite current reductions in training expenses.
The introduction of Claude 3.7 Sonnet reflects a significant cost reduction in training advanced AI models, yet forecasts indicate future developments may require far greater investments, highlighting the evolving landscape of AI training economics.