AI-NAV · Article
Chinese AI Models Lead Global API Usage: OpenRouter Top 10
The leaderboard that matters most in AI is not a benchmark — it is what developers actually call. OpenRouter, the API aggregator routing traffic for millions of developers, publishes exactly that. This week's numbers tell a story that would have been unthinkable two years ago: every one of the five most-used models on Earth is Chinese.
1. This week's top 10
Tokens processed through OpenRouter, week ending October 11, 2026.
| Rank | Model | Lab | Tokens | Week-over-week |
|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | DeepSeek (CN) | 41.0T | +60% |
| 2 | GLM 5.3 Flash | Zhipu (CN) | 12.2T | +26% |
| 3 | MiMo-V2.6-Flash | Xiaomi (CN) | 11.5T | +19% |
| 4 | Step 5 Preview | StepFun (CN) | 8.84T | new |
| 5 | Hy4 preview | Tencent (CN) | 8.28T | +27% |
| 6 | GPT-6 Luna | OpenAI (US) | 7.08T | +15% |
| 7 | Nemotron 3 Ultra (free) | NVIDIA (US) | 6.87T | +22% |
| 8 | Claude Opus 5.5 | Anthropic (US) | 5.02T | +104% |
| 9 | DeepSeek V4 Flash 0731 | DeepSeek (CN) | 4.92T | -18% |
| 10 | Space Bunny Alpha | Stealth (unknown) | 4.2T | -89% |
*Source: OpenRouter rankings (Oct 11, 2026), cross-checked with 每日经济新闻's weekly tally.*
2. The headline numbers behind the sweep
Zoom out from the top 10 and the pattern is starker. Total weekly volume on the platform reached 172 trillion tokens (+3.6%). Of the ranked models, Chinese labs processed 86.74 trillion (+51% week-over-week) against 18.97 trillion (+17%) for US labs — the 24th consecutive week Chinese models have led. DeepSeek V4.1 Flash alone handled 41 trillion tokens, more than double the entire US contingent combined.
Two side stories worth noting. The anonymous "Space Bunny Alpha" — a stealth model that briefly topped the chart last week — collapsed 89% and fell to tenth, a reminder that mysterious benchmark-hunting models burn bright and short. And Anthropic's Claude Opus 5.5 doubled its volume (+104%) within three weeks of launch, proving premium Western models can still grow fast when the price-performance math improves (it launched 20% cheaper than its predecessor).
3. Why Chinese models win the volume war
Three structural reasons, and none of them is a secret:
Price. DeepSeek V4.1 Flash lists at $0.15–0.30 per million input tokens and $0.60–1.20 per million output — against $2–10 for US mid-tier models and up to $50 for premium flagships. When the price gap is 10–50×, "good enough" wins the bulk traffic.
Open weights. DeepSeek, GLM and Qwen release weights, which spawns a reseller and self-hosting ecosystem that further drives down cost and friction. Intel CEO Lip-Bu Tan said it plainly this week: Chinese frontier models are being adopted rapidly because of low token costs and the efficiency of open-source development.
The routing reality. As our pricing guide documented, roughly 70% of real-world queries can be handled by budget models. Developers have done the math — and the budget tier is overwhelmingly Chinese. Notice that the top three are all "Flash"-class small models: volume lives in the cheap tier.
4. But volume is not revenue: two different battlegrounds
Before declaring a winner, be precise about what this chart measures. Token volume is not users, not revenue, and not profit. US frontier models charge 10–50× more per token, so a far smaller volume can still mean a larger business. OpenAI and Anthropic dominate consumer subscriptions ($20–200/month plans) and enterprise contracts — money that never touches OpenRouter.
The honest reading: the market has split. US labs win the premium and consumer-subscription battleground; Chinese labs win the high-volume API battleground. Claude Opus 5.5's +104% and GPT-6 Luna's #6 slot show the US is not absent from volume — it is just fighting with different weapons. And the new entrants keep coming: Step 5 Preview (a 600B MoE agentic model) debuted at #4 within days of listing, and Claude Haiku 5.5 and Mistral Large 4 just appeared on the trending chart.
5. What this means for you
Developers: the market has already voted for routing. Default bulk traffic to Chinese Flash-tier models, reserve frontier models for the hardest ~10% of tasks, and keep your provider abstraction layer warm — this leaderboard reshuffles monthly.
Tool users: cheap API volume flows downhill into cheaper apps. The AI tools you use daily are increasingly built on these models, whether or not the label says so.
Watchers: the thing to monitor is not whether Chinese models lead volume — they will, at these prices — but whether they convert volume into durable ecosystems: developer loyalty, enterprise contracts, and the agentic workloads that Step 5 Preview is explicitly built for.
6. Read this chart with three caveats
- OpenRouter is one channel. It skews toward developers and indies; direct API contracts and consumer apps (ChatGPT, Doubao, Qwen) are invisible here.
- Tokens are not quality. A trillion tokens of casual chat and a billion tokens of drug-discovery reasoning count the same.
- Stealth models distort the board. Space Bunny Alpha's rise and collapse in a fortnight shows how much noise anonymous models inject into weekly rankings.
Primary sources (accessed October 12–13, 2026): OpenRouter LLM rankings (weekly, monthly and trending views); 每日经济新闻 weekly OpenRouter tally (Oct 12, 2026); Reuters via AI Consultant brief on Intel CEO comments (Oct 12, 2026); StepFun Step 5 Preview announcement (Sep 2026); Anthropic Claude Opus 5.5 release notes (Sep 2026).
*Note: rankings change weekly. Figures are a snapshot of the week ending October 11, 2026 — check OpenRouter for the live board.*