Chinese open-weight models are winning US developers on price — and Washington is noticing

A wire story from the Associated Press, published July 26, 2026 and carried by ABC News, WRAL and other outlets, reports that Chinese AI models are being adopted by American companies and developers at a noticeable rate, with cost cited as the dominant reason.

The story's anchor example is Raffi Krikorian, chief technology officer at Mozilla, who told AP he switched much of his day-to-day work to Moonshot's Kimi K3 within days of its release, having previously used Z.ai's GLM-5.2 for routine tasks. AP also reports that Coinbase said it is moving to Chinese models to reduce costs. Both are reported claims from named sources, not independently established facts about relative model quality.

What is actually new here

The DeepSeek moment of early last year established that a Chinese lab could match US models cheaply. What the AP piece describes is a later stage: several Chinese labs shipping in quick succession — Z.ai's GLM-5.2 in June, Moonshot's K3 in July, an Alibaba Qwen3.8 Max preview in July, DeepSeek V4 previews in April — and US buyers routing production work to them.

Two adoption datapoints are offered. AP says the five most-used models on OpenRouter over the past month were Chinese, and cites Sensor Tower estimates of more than 930,000 Kimi downloads in the week following K3's release, up about 200% week over week, with roughly 86,000 US downloads, up 387%. AP also reports Moonshot temporarily suspended new subscriptions as demand pushed it near capacity.

Both figures deserve care. OpenRouter routes a particular slice of developer traffic and is not a proxy for total enterprise usage; app download counts measure consumer curiosity after a launch spike, not sustained use. Percentage jumps off a low base are the least informative numbers in the story. An editor may want to seek the underlying Sensor Tower methodology and a longer OpenRouter time series.

The economic argument being made

The clearest driver in the reporting is unit price. Curt Meinhold, a North Carolina technology executive quoted by AP, framed it as a "good enough" calculation — comparable performance on code and research at a fraction of the per-million-token output price. AP attributes to Goldman Sachs a July research note describing Chinese models as reaching a "critical stage" for wide adoption, and cites analysts including Alex Colville of the Australian Strategic Policy Institute on how agentic workloads compound per-token cost differences across many steps.

The Goldman note and its exact wording should be verified against the primary document before publication; the AP story characterizes it rather than quoting at length.

The limits the story itself flags

AP does not present this as a clean Chinese lead. Anastasios Angelopoulos of the evaluation platform Arena told the outlet that Chinese models still lag US leaders across full-range capability. Yasir Atalan of CSIS is cited saying US firms are also pushing out cheaper tiers of competitive models — meaning the price gap is a moving target, not a fixed structural advantage.

The financial picture is thinner than the adoption picture. AP reports Z.ai's revenue rose 132% to 724 million yuan (about $107 million) last year while net loss grew 60% to 4.7 billion yuan (about $694 million). That is a company losing roughly six-and-a-half times its revenue. Cheap inference that is subsidized is not the same thing as cheap inference that is sustainable, and the story does not resolve which this is.

Openness as strategy, and as policy problem

Most Chinese models are open-weight while leading US frontier models are closed, and AP quotes Omdia's Lian Jye Su saying Chinese vendors are expected to use openness to drive global adoption. Krikorian's line to AP — that "the open frontier is becoming increasingly Chinese-built" — is the sharpest framing in the piece, and it is one practitioner's assessment, not a measured finding.

On the policy side, AP reports the Trump administration on Wednesday accused Moonshot of using "covert" but not necessarily illegal methods to build K3 off Anthropic's Fable, part of a broader distillation dispute in which US politicians and companies including Anthropic have alleged extraction of their technologies and Beijing has called the claims "groundless." These are allegations; nothing in the reporting establishes them as adjudicated.

The story also notes a counterintuitive dynamic: AP says Z.ai released GLM-5.2 shortly after the administration placed export controls on Anthropic's Fable and Mythos models, keeping them offline for more than two weeks, with Angelopoulos observing that restricting an American model can create an immediate opening for a Chinese one. Separately, AP reports Microsoft, Meta and Nvidia signed an open letter published Friday backing open models — an item worth confirming directly against the letter, as the AP summary does not give its full text, signatory list or specific asks.