Chinese models took all top five spots among developers on OpenRouter

Chinese AI models held the top five spots on OpenRouter's weekly ranking for the week of July 28-August 3, 2026, according to Chinese state media outlet CGTN.

OpenRouter is a service that lets developers connect to many different AI models through a single access point. The ranking tallies how many tokens each model has processed. Tokens are the small chunks of text that AI models read and write, so the more tokens, the more usage.

Top five

According to CGTN, the list looked like this:

1. DeepSeek V4-Flash: 7.1 trillion tokens 2. Xiaomi's MiMo-V2.5: 5.4 trillion 3. Tencent's Hy3: 4.89 trillion 4. DeepSeek V4-Pro: 3.1 trillion 5. Z.AI's GLM-5.2: 2.87 trillion

DeepSeek thus had two of the top five models. Of roughly 56.8 trillion tokens in total that week, 28.13 trillion went to Chinese models and 4.38 trillion to American ones. According to the tally, China has led in token consumption for 14 weeks in a row, and nine of the ten most-used models were Chinese.

The previous week's top five was also all Chinese, but in a different order. Xiaomi MiMo-V2.5 was in first place with 10.5 trillion tokens, ahead of DeepSeek V4 Flash with 6.37 trillion.

Price plays a role

CGTN points to price. In a pricing test, DeepSeek V4-Flash cost 3 cents per test. That compares with 86 cents for Moonshot's Kimi K3, $1.86 for OpenAI's GPT-5.6 Sol and $3.15 for Anthropic's Claude Fable 5. On an intelligence index from Artificial Analysis, DeepSeek V4-Flash scored 50 points out of 100, the same as Google's Gemini 3.6 Flash. Z.AI's GLM-5.2 and Meta's Muse Spark 1.1 scored 51.

CGTN attributes DeepSeek's strength among developers to better inference (the actual running of the model), dynamic KV caching (a way of reusing intermediate calculations), traffic management and long-context optimization. That is the outlet's own assessment.

What the figures don't show

The ranking is not a measure of all AI usage in the world. According to 36Kr, it reflects paid usage by foreign, independent developers via OpenRouter. It does not include traffic from the Chinese internet giants' own apps, private enterprise deployments or the American giants' closed APIs.

High usage also says nothing about quality. A cheap model can rack up many tokens simply because it is cheap. The figures also come from a state media outlet and from 36Kr, and small discrepancies appear between the sources for the previous week.