Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Chinese models route four times US volume on OpenRouter as DeepSeek jumps 801%

Chinese AI models are handling roughly four times the routed volume of their US counterparts on OpenRouter, according to Dealroom data, with DeepSeek's usage up 801%. The numbers turn a developer preference into a visible market signal about price and switching costs.

Published
Dealroom:中国模型在OpenRouter的调用量约为美国模型四倍,DeepSeek涨801%
Image source: deepseek.com

Chinese AI models are handling roughly four times the routed volume of their US counterparts on OpenRouter, according to data published by Dealroom, with DeepSeek's usage up 801%.

OpenRouter functions as a router for model choice: applications call many vendors' models through one interface and pay per token. Share shifts there track production traffic more closely than vendor-reported benchmark scores do.

Price is the obvious mechanism. Chinese open-weight models have generally been priced below comparable closed models, and a router reduces switching to a configuration change, collapsing migration cost to near zero.

The caveat is scope. Router traffic skews toward third-party apps and individual developers, while enterprises running their own inference clusters stay invisible in these figures. Four times describes the observable slice of the market, not all AI usage.

Concentration cuts both ways. The same ease of switching that built these numbers can reverse them quickly if a model runs into reliability problems or regulatory scrutiny.

Read as a trend rather than a verdict, the data points to two questions: whether DeepSeek's growth holds up on dimensions other than price, and whether US vendors respond with cuts or with agent capabilities strong enough to pull the traffic back.

For now the scoreboard is denominated in tokens and invoices rather than leaderboard points. That is a harder test to game, and a more honest one for buyers deciding what to run in production.

Why it matters

Routing data pulls open-model competition back to cost and availability. Continued Chinese share gains would squeeze closed-model pricing power and developer lock-in.

DeepSeekOpenRouterMarket Data
Back to realtime news

Nearby Updates

All

10/27, 17:53

Lenovo unveils AI Host P7, a 300-gram device running a 125-billion-parameter model locally

Lenovo has announced the AI Host P7, a device reported at roughly 300 grams that can run a 125-billion-parameter model on-device, according to Pasquale Pillitteri. If the specification holds up, it marks another step in moving large-model inference from cloud servers onto personal hardware.

10/27, 17:57

OpenChiip open-sources a digital employee pitched as an AI counterpart for every engineer

According to 36Kr, OpenChiip has open-sourced its digital employee product, positioned as giving every engineer a personal AI counterpart. Public information so far stops at the launch itself: the underlying models, licence and concrete capabilities have not been disclosed.

10/27, 18:06

Haut.AI launches SkinFM skin foundation model alongside SkinBench benchmark

Haut.AI has released SkinFM, a foundation model built specifically for skin, together with a new benchmark called SkinBench whose results expose flaws in how AI systems currently analyze skin. The company, which sells AI skin analysis software to beauty brands, is pairing a vertical model with a public evaluation in a single push.

10/27, 16:51

Google Cloud adds OpenAI models to Model Garden, led by GPT-6 Astra

Google Cloud is now offering OpenAI models through Model Garden, placing three of OpenAI’s closed models alongside more than 200 Google, third-party and open models in one catalog. The first three are GPT-6 Astra, GPT 6.1 Sol and GPT 6.0 Luna, sold through Google Cloud Marketplace and billed on a single Google Cloud invoice.