Many developers choosing AI coding assistants have long dismissed Chinese-made models as mere imitations of US models at best, or outright plagiarism at worst. When Anthropic officially accused Chinese firms of a "distillation attack" in February 2026, that suspicion only deepened. Yet just a month later, US coding tool Cursor admitted that it had built its own model on top of an earlier-generation Kimi K2.5 without initially disclosing this fact—and four months after that, on July 16, Kimi K3, a 2.8-trillion-parameter model, topped Arena's frontend coding rankings with a score of 1679 (with reaction spreading through the 17th). Behind this rapid shift—from suspicion to open dependence on the very technology under scrutiny—lies an asymmetry that investment figures alone cannot explain.
A Design That Activates Only 16 of 2.8 Trillion Parameters
On July 16, 2026, Moonshot AI (based in Beijing, backed by Alibaba) unveiled Kimi K3, built on a Mixture of Experts (MoE) architecture. With 2.8 trillion total parameters, the company claims it as "the world's largest open-weight model." Of its 896 experts, only 16 are actually activated at any given time, using a proprietary design called Kimi Delta Attention and Attention Residuals. It supports a context length of 1 million tokens and features native multimodal capability, handling images and video directly.
Even an organization with 896 experts on staff doesn't need to convene all of them for a single task—it only needs to call in the 16 who actually know the subject. Kimi K3's MoE architecture follows the same logic. For any given inference, only a small fraction of the total model actually performs the computation, allowing it to hold the equivalent of 2.8 trillion parameters' worth of knowledge while running at a cost much closer to that of a far smaller model. This is precisely the mechanism that lets open-weight players keep serving costs low even as they pile on total parameters.
Benchmark results bore out this design. On the rankings of Frontend Code Arena, run by Arena, Kimi K3 took the top spot with a score of 1679, surpassing both Claude Fable 5 (1631) and GPT-5.6 Sol. Arena co-founder and CEO Anastasios Angelopoulos commented on X that this result was "perhaps the biggest release of the year." Russ Salakhutdinov, a Carnegie Mellon professor and former Apple AI research director, also posted that it was "a huge win for the open-source community."
The API is priced at $3 per million input tokens and $15 per million output tokens—roughly ¥450 and ¥2,250 respectively at ¥150 to the dollar. BofA analysts characterized this as "roughly half" the price of GPT-5.6 Sol ($5 input / $30 output). Doing the math, input comes out 40% cheaper and output exactly 50% cheaper—a clear price gap versus the US players.
Just Five Months After Distillation Allegations, Hidden Dependence Became an Open Top Ranking
The story goes back five months before K3's release. On February 23, 2026, Anthropic officially announced that three companies—DeepSeek, Moonshot, and MiniMax—had carried out a "distillation attack," systematically harvesting Claude's outputs through roughly 24,000 accounts across more than 16 million interactions. Of these, Moonshot alone reportedly accounted for over 3.4 million. It was one of the industry's most serious accusations: that Chinese firms were illicitly siphoning off the fruits of US models' work.
A month later, the situation took an unexpected turn. On March 22, 2026, TechCrunch reported that Anysphere, the company behind the US coding tool Cursor, had built its new coding model, Composer 2, on top of Moonshot's Kimi K2.5—without disclosing this fact anywhere in its initial announcement. Kimi's official account explained it as a "legitimate commercial partnership" conducted through Fireworks AI, and Cursor co-founder Aman Sanger acknowledged, "It was a mistake not to state from the outset in our blog post that it was Kimi-based." In other words, one of America's leading developer AI tool companies had already come to depend on a model produced by the very company accused of the distillation attack—and had kept that fact quiet. It was the moment a striking arrangement came to light: the accuser's and the accused's technologies had been coexisting, unannounced, within the very same development workflow.
Four months further on, Kimi's latest version, K3, took sole possession of the top spot in Arena's coding rankings. Technology that in February stood accused of plagiarism had, by March, become the quiet foundation US companies relied on—and by July, had turned into the open challenger seizing the top spot outright. This five-month arc is precisely the reversal that coverage focused solely on comparing prices and specs tends to miss.
On Artificial Analysis's GDPval-AA v2, K3 reportedly ranks third with a score of 1668, trailing both Claude Fable 5 (1760) and GPT-5.6 Sol (1748). Technology analyst Patrick Moorhead reportedly characterized the reaction to this event as "a strikingly similar overreaction to the DeepSeek shock of 2025." The top ranking is confined to a coding-specific benchmark, and it's far from a claim to absolute supremacy.
A 23x Investment Gap, a 2.7% Performance Gap: What a $273.5 Billion Difference Reveals
According to Stanford University's AI Index 2026 edition, private AI investment in 2025 totaled $285.9 billion in the US versus $12.4 billion in China. Subtracting $12.4 billion from $285.9 billion yields a difference of $273.5 billion, or roughly a 23-fold gap. Given a funding disparity of this magnitude, one would naturally expect a correspondingly large gap in performance.
And yet, the same research shows that as of March 2026, the gap between the top-performing US and Chinese models stood at just 2.7%. Placed side by side, a 23-fold investment gap and a 2.7% performance gap make for a jarringly poor match. This doesn't prove that investment fails to translate into performance, but at minimum, it's a number that shows Chinese players currently catching up to a degree far beyond what the investment gap alone would predict.
Kimi K3's pricing—¥450 input and ¥2,250 output per million tokens (at ¥150 to the dollar)—comes in at roughly half of GPT-5.6 Sol's. The market has already begun reacting to this price gap. One observation notes that during a single week in February 2026, Chinese-made open-weight models accounted for roughly 61% of tokens processed among the top 10 models on OpenRouter, with Chinese models occupying four of the top five spots. That said, this is a momentary figure limited to a specific week and the top 10 models—not a full-period market share figure across the 400-plus models available. This price asymmetry mirrors the same mismatched proportions seen between the investment gap and the performance gap.
A similar shock rippled through the industry with the DeepSeek-R1 shock of January 2025, and US players reportedly reclaimed the benchmark lead within a few months afterward. Whether this instance repeats that pattern, or instead plays out as a case—like Cursor's—where dependence becomes an established fact before any catch-up occurs, will shape how it's ultimately understood. This is not the first time this dynamic has played out.
Japanese Companies Weigh Half-Price Costs Against Geopolitical Risk
Kimi K3's pricing of ¥450 input and ¥2,250 output puts it within reach as a realistic option for cost-conscious development teams in Japan. At the same time, for government-related projects or contracts involving security review, the country of origin of a model can itself affect procurement requirements. Choosing a development tool now means weighing this geopolitical risk alongside performance and price—precisely the choice Cursor faced without disclosing it. The temporary share figures observed on OpenRouter are not irrelevant to Japanese companies either.
Around the same time, at the World Artificial Intelligence Conference (WAIC) that opened in Shanghai, Huawei unveiled the Atlas 950 SuperPoD, a configuration capable of connecting up to 8,192 Ascend NPUs, claiming 6.7 times the computational power of NVIDIA's NVL144. In his speech, President Xi Jinping stated that "AI should not be a solo performance by a single nation, but a symphony of international cooperation," and 29 countries signed on to establish the World AI Cooperation Organization (WAICO). While Cursor's dependence had been limited to model weights, Huawei's exhibit points to the next stage: dependence on the computing infrastructure itself. Amid ongoing debate over export controls targeting China, the fact that both the top-ranked open-weight coding model and the underlying computing infrastructure originate from China creates a gap between regulatory intent and on-the-ground reality.
Even Cursor, one of America's leading development tool companies, ended up explaining what was inside its own product only after the fact. There is no guarantee that performance charts and benchmark scores can be trusted unconditionally at face value, and Japanese companies too now face the challenge of asking not just "which model performs best," but "what is actually powering this tool."
The July 27 Weight Release Will Test Whether This Reversal Is Fleeting or Structural
In the five months between February's accusations and July's top ranking, a Chinese-made model shifted from "object of suspicion" to "openly relied-upon technology." Following the precedent of the DeepSeek-R1 shock, US players may well reclaim the benchmark lead within the coming months. But as the Cursor episode shows, even after rankings shift, the underlying dependence itself tends to persist.
Moonshot has stated it plans to release the model's complete weights by July 27. If this happens, outside researchers will be able to directly verify the 16-of-896-expert design and the million-token context handling, revealing whether K3's performance holds up to its claims. Regardless of which side ultimately holds the top spot, how openly US development tool companies come to acknowledge their reliance on Chinese-made open-weight models will be the next fork in the road—determining whether this five-month reversal turns out to be a passing moment or a new premise for the industry.