“The DeepSeek model can be run in production for approximately 20% of the cost of running models from OpenAI or Anthropic, with comparable or better latency and reliability.”