89% of 4%: How strong China's open AI models really are

What's going on

Two numbers from the same week paint opposite pictures of China's open AI models. The first comes from venture capital firm Menlo Ventures: of the money US companies spend on open models, 89% goes to Chinese providers. Moonshot AI's Kimi alone accounts for 61%, DeepSeek for 15% and Zhipu's GLM for 11%. Chinese outlets like The Paper are reporting this as a breakthrough with Western enterprises.

The second number comes from benchmarking service Artificial Analysis. Since its index switched to version 5 on October 27, the best model with released weights sits 26 points behind the top. As recently as September, a Mozilla report had put the gap at around four months. The debate is already underway in Vietnam and China. VnExpress writes that China's open models might be “weaker than thought”. On Tencent News, the question is whether they've been “knocked back to their true form”.

Our thesis: both numbers are correct, but they describe different things. China dominates the low-cost market for open models by a wide margin. But part of their apparent closeness to the frontier seems to have rested on public tests the models may have already been familiar with.

89 percent – of a small pie

Menlo's report “2026: The State of Enterprise AI” is a survey, not a billing analysis. Morning Consult surveyed just under 400 US decision-makers between September 8 and 26, 2026, who together control more than $4 billion in AI budget. Respondents stated what share of their AI workload runs on which model. Menlo weighted these answers by spending and cross-checked them against token prices from OpenRouter, Ramp and Vercel. Tokens are the text units models are billed by. The market shares are therefore extrapolated estimates.

We did the math. Open models process 19% of tokens at $0.13, closed models 81% at $0.70. That gives a spending share of roughly 4.2%. So the 4% follows almost entirely from the price difference of about 5.4 to 1. It's not an independent second measurement. The 89% is therefore 89% of just over four percent. For comparison: according to the same report, Anthropic accounts for 55% of all spending on model APIs, OpenAI for 31%.

There are also potential conflicts of interest. Menlo is an Anthropic investor. According to the report, Fireworks and Modal are Menlo portfolio companies, as is Cursor, and a large share of open-model volume runs through these inference clouds. FourWeekMBA also lists OpenRouter as a Menlo investment.

Why Kimi, of all models?

The 61% for Kimi doesn't match what you can see publicly. In the current weekly ranking on OpenRouter, a routing service that forwards and bills requests to many models, Kimi doesn't appear among the ten most-used models. The top spots go to two DeepSeek variants with 56.5 and 16.5 trillion tokens, followed by GLM 5.3 Flash with 15 trillion.

Two explanations suggest themselves. First, Kimi is significantly more expensive per token than DeepSeek. Count by money instead of tokens, and the shares automatically shift toward Kimi. Second, Menlo itself mentions that Cursor built its coding model Composer 2 on Kimi K2.5, with Fireworks supplying the compute for training. Cursor is one of the highest-grossing AI tools, with estimates putting its annual revenue at $2 to 3 billion. If surveyed companies logged their Composer usage as “Kimi”, a single product could push the figure up significantly. Our take: that's plausible but unproven. Menlo doesn't break down the Kimi share.

Moonshot is genuinely growing, too. According to Bloomberg reports, annual revenue rose from around $200 million in April to over $1 billion in August. The target is $2 billion by year-end. These figures were reported by the company itself and haven't been audited.

26 points – and why it's now more like 12 to 14

Artificial Analysis now builds 60% of its index on private test sets or withheld answers. In v4.3, that share was 45%, according to secondary reports. The goal is to prevent contamination: a model shouldn't score well because test questions and their solutions were in its training data. At the same time, Artificial Analysis changed the weighting. Instead of 30/30/20/20, general knowledge, knowledge work, coding and science now count 25% each. New additions are FrontierCode 2.0 (real-world programming) and Terminal-Bench-Science (long scientific workflows in the terminal).

According to Artificial Analysis, this widened the gap between the best open model and the top from 3 to 12 points. The 26 points only apply to models whose weights were already downloadable on October 27. That caveat is now outdated: Moonshot posted the weights of Kimi K3.1 Pro on Hugging Face on November 4. On the November 6 leaderboard, Gemini 4 Carbon leads with 60 points, while Kimi sits at 46. So the gap for the best downloadable models is 12 to 14 points. That's far less than 26, but four times as much as under v4.3.

The individual tests are revealing. On Terminal-Bench-Science, GPT-6 Astra scores 63%, Kimi 34% and DeepSeek V4.1 Flash 13%. On HLE-Diamond, it's 68% for Opus vs. 37% for Kimi. On the private FrontierCode 2.0, however, DeepSeek reaches 28% vs. 46% for Opus. So the gap in programming is clearly smaller than in science.

Four months or 26 points? Different yardsticks

Mozilla's September 15 report is based on data from the research group METR. It measures how long tasks can be (in human working hours) that a model still solves half the time. The result: open models reach the level of closed models about 4.4 months later. On the old Artificial Analysis index, Kimi K3 was only 3 points behind Claude Opus 5. Epoch AI arrived at about four months with its Capabilities Index in May. An Epoch analysis cited by Startup Fortune puts the average for Chinese models since 2023 at seven months, ranging from four to fourteen. Mozilla itself advocates for open models. Its CTO called the report partly advocacy.

Points and months can't be converted directly into each other. More importantly, v5 changed two things at once: more private tests and more weight on science. Our take: from the outside, you can't separate how much of the drop is due to contamination and how much to the new weighting. The “memorization” thesis is therefore supported, but not proven.

Nathan Lambert, who covers open models on his blog Interconnects, defends both sides:

“The previous index flattered the open models too much.”

At the same time, he considers the end of open models “greatly exaggerated”. The best labs “benchmax a little”, meaning they optimize specifically for test scores; the worst do it “a lot”. He criticizes that private tests take away the public's ability to find errors and that Artificial Analysis is paid by the labs it evaluates. He expects leading open labs to catch up significantly within two to three months. Dean Ball, now at OpenAI according to Weijin Research, said by contrast that the new leaderboard matches his hands-on impression for the first time.

OpenRouter: 58, 46 or 30 percent?

OpenRouter's usage figures also diverge because they measure different things. In July, Bloomberg reported a record share of 58%: that's how much of the tokens from accounts with a US billing address ran through Chinese models, briefly even 63% in the first week of July. A few weeks earlier, CNBC had reported a peak of 46% and “over 30%” in every week since February 8. The frequently cited 30%, on the other hand, comes from a different Bloomberg chart. It shows the share of US models across the entire platform, which fell from 70% (June 2025) to 30% (June 2026).

The outlet Noise Cancelling points out the limits of this data. Customers of a routing service are especially quick to switch to the cheapest model. Over half of the open-model traffic there is roleplay and creative writing. And Anthropic accounts for only 12 to 15% of tokens on OpenRouter but gets about half the money.

Bottleneck: the data center

If you want to dominate the low-cost market, you need compute. That's where things fall short. According to Reuters, around 600,000 Nvidia H200 chips are stored in Hong Kong because Beijing isn't approving large imports. On top of that, inference, meaning the day-to-day operation of the models, is supposed to run on Chinese chips. According to CNBC, Moonshot is negotiating for data centers in Europe with 8,000 or more Blackwell GPUs, mainly for inference. After launching Kimi K3 in July, Moonshot temporarily had to pause new sign-ups. Tencent increased capacity tenfold for Hy4, released on November 4, after it was overloaded on day one.

Our assessment

Our take: China's labs are clearly winning the open-model segment. Whether you look at the US survey, OpenRouter or downloads, Western competitors like Llama barely matter anymore. But this segment accounts for only a small share of spending, and the 89% rests on survey estimates. With the Kimi figure, it's unclear how much of it hinges on a single customer.

On performance, the gap to the frontier is larger than the four-month headlines suggested. It's especially large in science and smaller in programming. The 26 points were a snapshot taken before Kimi K3.1 Pro's release. What remains open is whether the private tests mainly expose contamination or simply reward different capabilities. For you as a user, that means: if you want to write code or process large volumes of text cheaply, Chinese models give you a lot of performance for your money. If you're working at the edge of what's possible, say in research, you'll still end up with OpenAI, Anthropic or Google.