A version of Chinese startup DeepSeek's flagship AI model is by ‌far the least expensive to run on benchmark tests among well-known models globally and more than 100 times cheaper to run than Anthropic's Claude Fable 5, according to a research firm.