When Bloomberg Businessweek and testing firm Vals.ai asked a lineup of US and Chinese AI models to build the same fictional coffee e-commerce site, nearly all of them delivered a working product. The difference was the bill. Anthropic’s Claude Fable 5 rang up $48.99 for the task. DeepSeek’s V4-Pro? Under four bucks.
The July 2026 test, centered on building a site called Brewberg, wasn’t designed to crown one model the best coder. It was designed to answer a simpler, more consequential question: if most frontier models can do the job, why would anyone pay ten times more for one over another?
The price gap is enormous
Anthropic’s Fable 5 charged roughly $50 per million output tokens. Moonshot’s Kimi K3 came in at $15 per million tokens. DeepSeek’s V4-Pro sat at $3.96 per million tokens, making it roughly 92% cheaper than Anthropic’s offering for comparable coding output.
A broader Bloomberg analysis from August 2026 confirmed this wasn’t a fluke of one test. Chinese models were consistently 60-90% cheaper than their American counterparts across a range of tasks, depending on the specific model pairing.
