Anthropic AI Training Costs: Was it Cheaper Than Expected?

Claude 3.7 Sonnet Training Costs Revealed

Anthropic’s recently launched Claude 3.7 Sonnet AI model was trained at a cost of “a few tens of millions of dollars.” This figure represents the computational expense, utilizing less than 10²⁶ FLOPs of processing power.

This information was shared by Wharton professor Ethan Mollick via a post on X (formerly Twitter). He indicated that Anthropic’s PR team clarified that Sonnet 3.7 does not qualify as a 10²⁶ FLOP model.

Confirmation and Further Development

TechCrunch contacted Anthropic to verify these details, but a response was not received prior to this article’s publication.

Despite this, the reported training cost suggests a trend towards more affordable development of cutting-edge AI models. Future iterations, however, are anticipated to require significantly greater investment.

Cost Comparison with Previous Models

The training expenditure for Claude 3.7 Sonnet aligns with the cost of its predecessor, Claude 3.5. Anthropic CEO Dario Amodei previously disclosed that Claude 3.5 also required a few tens of millions of dollars for training.

These costs are notably lower when contrasted with the expenses incurred by other leading AI developers in 2023. OpenAI’s GPT-4 reportedly exceeded $100 million in training costs, as stated by CEO Sam Altman.

Gemini Ultra's Expense

Furthermore, a Stanford study estimated that Google invested approximately $200 million to train its Gemini Ultra model.

Future Trends in AI Training Costs

While current models demonstrate decreasing training costs, Anthropic’s Dario Amodei predicts that future AI models will likely require investments reaching into the billions of dollars.

It’s important to note that training costs represent only a portion of the overall expense. Significant resources are also dedicated to safety testing and fundamental research.

Rising Operational Costs

As the AI industry shifts towards “reasoning” models – those capable of sustained problem-solving – the costs associated with running these models are also expected to increase due to extended computational demands.