🔥36氪•Stalecollected in 3h
China AI Tokens Surpass US, 4 Models Top Global 5
💡China LLMs #1 globally—test for cost/speed edge in your stack.
⚡ 30-Second TL;DR
What Changed
Chinese models: 4.12T tokens > US 2.94T (Sep 9-15)
Why It Matters
Signals China's lead in AI model adoption and affordability, challenging US dominance. Practitioners gain access to high-performing, low-cost alternatives via global APIs.
What To Do Next
Browse OpenRouter rankings and test top Chinese models for cheaper, high-volume inference.
Who should care:Developers & AI Engineers
🧠 Deep Insight
Web-grounded analysis with 6 cited sources.
🔑 Enhanced Key Takeaways
- •MiniMax M2.5 leads OpenRouter leaderboard with 4.26T tokens, followed by Kimi K2.5 at 3.94T, occupying top two spots[1].
- •Chinese models offer 76-99% lower costs than US premium models, e.g., Kimi K2 at $1.07 per million tokens vs Google Gemini 3 Pro at $4.50[2].
- •Chinese AI chatbots are free for consumers with zero subscription fees, driving 515 million users in China vs 195 million in US[2].
- •DeepSeek R1 trained for $294,000 at GPT-4 Turbo level, 99.7% cheaper than GPT-4, with 671B parameters (37B active)[2].
- •US leads in frontier capabilities by 7 months average since 2023 per Epoch Capabilities Index[4].
📊 Competitor Analysis▸ Show
| Model | Cost per Million Tokens (Blended) | Intelligence/Performance | Developer Country |
|---|---|---|---|
| Kimi K2 Thinking | $1.07 | Matches GPT-5 coding | China [2] |
| MiniMax M2 | $2.50 | Top OpenRouter rank | China [2] |
| DeepSeek V3 | Low (60-80% < GPT-4) | GPT-4 level | China [2] |
| Qwen-Plus | $0.30 (est.) | Open | China [2] |
| Google Gemini 3 Pro | $4.50 | 1490 Arena score | USA [2][6] |
| GPT-5 Codex | $15 output | Top-tier | USA [2] |
| Claude Opus 4.5 | N/A | 1469 Arena score | USA [6] |
🛠️ Technical Deep Dive
- •DeepSeek R1: 671B parameters (37B active), trained for $294,000, achieves GPT-4 Turbo performance level[2].
- •Kimi K2: Training cost $4.6 million, matches GPT-5 coding capabilities, parameters not disclosed[2].
- •Chinese models often use distillation from more capable (likely US) models to boost competitiveness[5].
🔮 Future ImplicationsAI analysis grounded in cited sources
Chinese models will capture 60%+ of global high-volume inference by 2030
Extreme price advantages (1/4 to 1/8 US levels) drive structural migration to high-frequency agent execution, expanding total AI market[1].
US revenue from enterprise AI will exceed $280B by 2030
US models dominate high-ARPU knowledge work with $14B (Anthropic) and $22B+ (OpenAI) annualized revenue from enterprise stickiness[1].
⏳ Timeline
2023-01
Epoch Capabilities Index tracking begins; US frontier models lead China by up to 14 months[4]
2024-05
First Chinese model surpasses GPT-4 ECI after 14-month gap[4]
2025-04
Baidu Ernie 4.0 made free for consumers following DeepSeek disruption[3]
2025-04
OpenAI o3 released; no Chinese model yet surpasses its ECI[4]
2026-02
OpenRouter data shows Chinese models at 4.12T tokens vs US 2.94T (Sep 9-15)[article]
2026-02
MiniMax M2.5 and Kimi K2.5 top OpenRouter leaderboard with 4.26T and 3.94T tokens[1]
📎 Sources (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 36氪 ↗