China's free open-weight models are turning the AI race into a price war

The competitive question in artificial intelligence has quietly changed. For three years it was whether anyone could match the capability of the leading American laboratories. This month it became something far more awkward for them: whether capability is worth paying for at all.

Google is months behind its own schedule for Gemini 3.5. According to Bloomberg, the company is holding the release while it tries to improve the model's coding performance enough to compete with Anthropic's Claude, widely regarded as the strongest coding assistant available. OpenAI shipped a new model at the start of July. Google had been expected to do the same in May. News of the delay pushed its shares down.

Into that gap stepped Moonshot. The Chinese startup released Kimi K3, a model that testers found unexpectedly strong at coding — particularly front-end work. "This may be the single biggest release of the year," said Anastasios Angelopoulos, chief executive of the AI testing company Arena. Within 48 hours Moonshot was rationing access. "Kimi K3 has received far more love than we expected, and our GPUs are feeling it," the company posted on X.

The detail that matters is not the benchmark score. It is that Kimi K3 is free to use and open-weight: anyone can download the core components and run the model on their own hardware. ChatGPT, Claude and Gemini are none of those things.

That is a deliberate national strategy rather than a single company's marketing decision. Chinese laboratories have consistently released open-weight models priced to undercut American rivals, mirroring the broader trade contest between the two countries. DeepSeek triggered a US market panic last year on the same logic. Alibaba previewed Qwen3.8 Max on Sunday with a parameter count in the same range as Kimi K3.

Beijing underwrites the approach directly, subsidising computing resources and power consumption for domestic AI startups so they can give their models away. The reasoning is straightforward. If Chinese firms cannot yet win on raw capability, they can win on cost — and their American competitors, unlike them, eventually have to turn a profit.

For Silicon Valley the consequence is structural. There are no long-standing incumbents in this market that customers default to out of habit. If businesses generating large volumes of code can get adequate results free, the case for paying premium prices for a brand-name model weakens with every release.

Washington's response is fragmenting rather than consolidating. Hawkish officials in the Trump administration are reviving efforts to block US access to Chinese models through commerce department rules or an executive order, Axios reported on Monday. Dean Ball, OpenAI's policy head for advanced AI and a former Trump adviser, wrote that Beijing's open-weight strategy pointed towards "full AI communism" — state datacentres supplying AI as a public good — and predicted the administration would eventually "create large amounts of regulatory risk around the use of open-weight Chinese models".

The government's own technologists are not persuaded. Emil Michael, the Pentagon's chief technology officer, replied publicly: "Every industry/ecosystem has its supreme village idiot. Dean Ball is that for AI." Meanwhile the White House, which until recently promised near-total deregulation, is restricting even allied access to advanced American cybersecurity models. Beijing is presenting a single position. Washington is presenting several at once.