One year ago, Chinese AI models had less than 2% of global API traffic. Today they own 61%. That’s not a market shift. That’s a heist — and nobody set off the alarm.
While Silicon Valley was busy arguing about AGI timelines on podcasts, a fleet of Chinese AI labs — DeepSeek, Kimi, MiniMax, Qwen, GLM — quietly became the engine powering the world’s software. Developers voted with their API keys. And they voted overwhelmingly for “same quality, one-thirtieth the price.”
The Numbers That Should Make US Labs Nervous
Here’s what happened on OpenRouter, the largest neutral AI traffic router in the world — the place where developers go when they want to pick the best model for each task without being locked into one provider:
Twelve months ago, US labs (OpenAI, Anthropic, Google) owned 70% of all token traffic. Today, they own about 30%. The remaining 70% has been swallowed by Chinese models. Xiaomi’s MiMo alone processes 4.21 trillion tokens per week — more than OpenAI’s entire API share. Trillion. With a T.
The price math is embarrassing for the Americans. DeepSeek V3.2 costs $0.28 per million input tokens. GPT-5.2 costs roughly $10 per million. That’s a 35x price difference for comparable — and sometimes better — coding performance. One San Francisco AI company, Lindy, switched from Anthropic’s Claude to DeepSeek and reported saving millions of dollars annually. A coding session that cost $10 on Claude? Fifty cents on DeepSeek.
Let that land for a second. Fifty. Cents.
The Part Where It Gets Complicated
It’s not just about price anymore. Kimi K2.5 from Moonshot AI scores 76.8% on SWE-bench Verified — the gold standard for real-world software engineering. That puts it ahead of most US models on the benchmark that actually matters to developers building production systems.
DeepSeek V4-Pro sits at 80.6% on SWE-bench. Opus 4.8 scores ~88.5% — so yes, the absolute top-tier US models still lead. But the gap has shrunk from a canyon to a crack, while the price gap remains a canyon.
Think of it like this: US labs built a Ferrari. Chinese labs built a Honda Civic that does 0-to-60 in 3.8 seconds. Most developers don’t need the Ferrari. They need to ship code by Friday.
Demis Hassabis — CEO of Google DeepMind, a man who does not engage in hyperbole — said at Davos that Chinese AI companies are now just six months behind the most advanced Western labs. Six months. That’s down from “several years behind” in 2024. The gap is closing faster than anyone publicly predicted.
What the Internet Is Saying
Of course, the tech world had thoughts:
Dario Amodei (Anthropic CEO): Chinese AI adoption isn’t just a business risk — it’s a national security question. He’s warned that feeding proprietary code and customer data into systems governed by Chinese law is “a legal exposure decision, not just a cost decision.”
Demis Hassabis (DeepMind CEO): “Six months behind the frontier.” Said at Davos with a straight face. The room got very quiet.
Andrej Karpathy (now at Anthropic): Predicted a “slopacolypse” in 2026 — a flood of low-quality AI-generated content overwhelming the internet. Relevant, because at 35x cheaper, you can generate approximately 35x more slop. Math checks out.
Developer community on HackerNews: “I don’t care where the tokens come from. I care that my startup has runway left.” The general vibe is pragmatic to a fault — and slightly anxious about what they just pasted into a Chinese API call.
Hot Take
Unpopular opinion: the US didn’t lose the AI model race — it won and then priced itself out of its own win. OpenAI built something extraordinary and then charged $10 per million tokens for it while DeepSeek said “here, same thing, fifty cents.” History will record this as the moment when the best product lost because it forgot that most customers don’t live in a Series B pitch deck. The real AGI was the hubris we made along the way.
This post has been created by Claude AI.
References
- Chinese AI Models Now Own 60% of Global API Traffic — ChatForest
- China’s Open-Weight Takeover — Data Gravity
- Chinese AI Models Lead OpenRouter Traffic: Coding Gains Come With China Data Risk — TechTimes
- The Open Weight Models that Matter: June 2026 — OpenRouter Blog
- AI luminaries at Davos clash over how close human-level intelligence really is — Fortune