Cheap AI Models Just Got Expensive

For two years, which AI model to route a workload through was an engineering call made on cost and quality. This week both inputs went to extremes at once. DeepSeek cut its V4-Flash pricing 50 percent on Saturday, one day after OpenAI cut its own prices by up to 80 percent, and according to independent benchmarking the same test suite now costs roughly 3 cents on DeepSeek's cheapest model against about 1.86 US dollars on OpenAI's and 3.15 on Anthropic's top model: a spread of two orders of magnitude, in a race Beijing is openly subsidizing even while warning its own firms about it. Then Congress showed what waits at the cheap end of that spread. Two House committees sent DoorDash's CEO a letter after the company's co-founder disclosed that DoorDash routes easier engineering tasks through Moonshot AI's Kimi model to cut costs, reserving Anthropic's models for the hard ones. That is exactly the optimization every competent engineering team is running right now. DoorDash owes Washington a complete list of every Chinese AI model it uses, plus security-testing records, by August 14, and in-person staff briefings by August 21. Stephen Forte on the structural forces underneath the cheap prices (a 20,000-chip Nvidia cluster reportedly provisioned to Moonshot through Alibaba, and a White House framework quietly finalized for the US labs), why the model-routing decision has left the engineering department, and the number on your cost dashboard that stopped telling the whole truth this week.