Photo: "Quantum Computing for Google Goggles" by jurvetson, BY via Openverse

Goldman Sachs Says Chinese AI Has Hit a 'Critical Stage' — the Math Explains Why

Per-token pricing gaps between U.S. and Chinese AI models used to be a rounding error for casual chatbot use — but multiply that gap by the hundreds of API calls a single autonomous agent makes to complete one task, and the math turns punishing for anyone still paying premium rates.

Goldman Sachs called this convergence a "critical stage" for Chinese AI adoption in a July research report, tying it directly to the global rise of "agentic" AI — systems that autonomously chain together multistep tasks rather than answering one prompt at a time. That distinction matters because agentic workflows don't buy AI once per task; they buy it dozens or hundreds of times per task, once for every step, tool call, and self-check the agent runs. Since providers bill per million tokens, a cost difference that looks incremental in a single query compounds with every additional agent step — according to analysts including Alex Colville of the Australian Strategic Policy Institute, that's what turns a linear price gap into an exponential one at scale. Curt Meinhold, founder of the digital legacy platform LilyList, put the arithmetic in plain terms: paying "a handful of cents per million output tokens" against Anthropic rates of "30 bucks or 40 bucks or 50 bucks" makes the decision automatic once a model is "good enough" — his term for where he says Chinese models like DeepSeek now sit for tasks like sales lead generation.

The adoption data backs up the shift in stated preference. Mozilla Chief Technology Officer Raffi Krikorian said he moved many of his daily workflows to Moonshot's newly launched Kimi K3 within days of its release, describing it to reporters as "snappier" than Anthropic's pricier Claude Fable; he said he'd already been running Z.ai's GLM-5.2 for routine calendar, document, and email management before that. Coinbase said it is shifting some workloads to Chinese models specifically to cut costs. Kimi's download numbers reflect the same swing: Sensor Tower estimated K3 pulled in more than 930,000 downloads in its first week — up 200% week over week, with U.S. downloads alone jumping 387% — a surge large enough that Moonshot said it had to temporarily cap new subscriptions.

None of this makes the Chinese models categorically better. Anastasios Angelopoulos, CEO of the model-evaluation platform Arena, said Chinese systems still trail U.S. frontier models on overall, full-range capability, even as they close the gap on specific tasks like code and research. And the compounding-cost logic cuts in an unexpected policy direction: when the Trump administration's export controls took Anthropic's Fable and Mythos models offline for more than two weeks earlier this year, Z.ai released GLM-5.2 into that exact window — prompting Angelopoulos to note that "restricting an American model can immediately create an opening for a Chinese competitor." The administration has separately accused Moonshot of using "covert" methods to build K3 off Anthropic's technology, without alleging the methods were illegal, while Beijing has called similar "distillation" claims from U.S. companies "groundless."

The open-source structure of most Chinese models adds a second multiplier on top of the pricing one: because competitors and independent developers can inspect and build on the underlying code, Mozilla's Krikorian argues "the open frontier is becoming increasingly Chinese-built" — a dynamic serious enough that Microsoft, Meta, and Nvidia signed an open letter this month backing open-weight AI models generally, even as Washington weighs further restrictions on the Chinese systems driving that openness.