A Chinese Lab Releases a Free Program That Reasons Like OpenAI's Best — and Says It Cost a Fraction to Build

DeepSeek's R1 shows its step-by-step thinking, matches OpenAI's o1 on math and coding tests, and can be downloaded by anyone. The company says it trained the underlying model for under $6 million, a claim that unsettled assumptions about who can build frontier AI and at what cost.

A Chinese company called DeepSeek released a program today called R1 that reasons through problems step by step, the way OpenAI's o1 does, and matches it on many math, science, and coding tests. Two things make it notable. It is free — anyone can download it and run it — and DeepSeek says the model underneath it was trained for under $6 million, a small fraction of what the American labs are believed to spend.

DeepSeek is an offshoot of a Chinese hedge fund, and it was not on most people's radar a month ago. The $6 million figure is disputed — it likely counts only the final training run and not the chips, the research, or the earlier models — but even allowing for that, the program was clearly built for far less than assumed, using fewer and less advanced chips than American companies rely on, partly because U.S. export controls keep the best chips out of China. That it shows its reasoning, in plain language, also makes it unusually useful to study.

The implications are still being worked out, and they cut in several directions. If a frontier-level reasoning program can be built cheaply and given away, the enormous valuations of the American labs — built on the assumption that this is expensive and scarce — look shakier. If China can do this under export controls, the controls may be working less well than intended. And if the best programs are free, everyone gets them, including whoever wants to misuse them.

The American reaction over the coming days is likely to be loud. For today, the fact is simple: the most talked-about AI program in the world this week is Chinese, open, and cheap.