There is a number working its way onto strategy decks this autumn, and it is blunt enough to survive the trip to a boardroom: roughly thirty to one. That is the gap between what the leading American AI labs charge per million tokens and what the best Chinese open-weight models now cost to run. Whether that number should change how a company buys AI is the argument of the moment.
The releases behind it came in a tight mid-2026 cluster. Four Chinese labs, Zhipu, MiniMax, Moonshot and DeepSeek, each shipped open-weight models that match or come close to the closed frontier on independent benchmarks for coding, tool use and reasoning. On one widely cited ranking, Z.ai's GLM-5.3 is the current pick for agent work and Moonshot's Kimi for coding. The pricing is where it bites. DeepSeek's own hosted interface runs at around $0.30 per million input tokens; Anthropic's and OpenAI's flagship models list at $10 or more. Because the weights are open, there is also a second door the Western labs do not offer: run the model yourself, on your own hardware, with no per-token fee at all.
The careful version of the claim, which is the one worth taking seriously, is not that the Chinese models are better. On the hardest reasoning and multimodal tasks the closed American models still lead, and the people making this argument say so plainly. The point is narrower and harder to dismiss: for the bulk of ordinary enterprise work, customer service, document summarisation, structured extraction, routine code, the quality gap between the best open model and the closed premium option is now smaller than the price gap. When that is true, paying the premium stops being a technical decision and becomes a procurement one.
One analyst gave the sharpest reading of why. What the American labs still have that DeepSeek and Kimi lack is distribution: sales channels, compliance certifications, security reviews, the procurement relationships regulated buyers need. But a distribution advantage, the argument goes, is cheaper to erode than a technology one, and it erodes on a timescale measured in quarters. On that view, a meaningful slice of today's frontier pricing is a distribution premium wearing the costume of a technology premium.
The same week sharpened the geopolitics underneath the economics. DeepSeek released software built to run on Huawei's Ascend AI chips rather than Nvidia's, described as open-source tooling to help Chinese silicon stand in for the hardware that export controls increasingly keep out. The cost and performance trade-offs against Nvidia are not yet measured and should not be assumed. But the direction is a full domestic stack, Chinese models on Chinese chips, offered to the rest of the world at a fraction of the Western price.
None of this means ripping out existing contracts. The sober recommendation is a hybrid: route high-volume routine work to the cheap open pathway, keep the expensive frontier APIs for the smaller set of genuinely hard, high-value tasks, and actually measure which workloads fall where. There are real reasons a buyer might still pay up, data-governance guarantees, support, and the underrated cost of running your own inference well. But the uncomfortable question the thirty-to-one number forces is no longer "which lab is best." It is "how much of what we currently pay is for capability, and how much is for a convenience we have stopped questioning?"