New AI model trained for just $249,000

Technology Sep 22, 2025

Chinese AI lab DeepSeek says it trained its new R1 model for only $249,000—a fraction of the $5.6 million spent on its earlier V3 model and way less than the estimated $100 million spent on models like GPT-4 at OpenAI.
The announcement came on Wednesday and is seen as a major disruption in the AI world.

Experts question if low price covers all development

Experts are curious if this low price covers all development or just the final training step.
DeepSeek is offering R1 at just $0.14 per million tokens (compared to OpenAI's $7.50), thanks to using 512 NVIDIA H800 chips—a more affordable, China-specific option—despite US restrictions on advanced chips.
With global AI spending expected to hit $1.5 trillion by late 2025, DeepSeek's approach could shake up how much future models cost.