In the latest LLM usage rankings from model aggregation platform OpenRouter, Chinese AI models have staged a historic takeover. DeepSeek V4 Flash, released on April 24, shot straight to first place in its debut month with 7.99 trillion tokens consumed.
Here are the top 10:
- DeepSeek V4 Flash: 7.99T tokens
- Hy3 preview (Tencent Hunyuan 3.0 preview): 7T tokens
- Claude Sonnet 4.6: 6.65T tokens
- Claude Opus 4.7: 6.07T tokens
- Hy3 preview (free): 5.83T tokens
- Kimi K2.6: 5.45T tokens
- Gemini 3 Flash Preview: 4.57T tokens
- DeepSeek V3.2: 4.06T tokens
- DeepSeek V4 Pro: 3.4T tokens
- MiniMax M2.7: 2.95T tokens
Beyond DeepSeek's three slots (V4 Flash, V3.2, V4 Pro), Tencent Hunyuan (Hy3 preview and free versions combined total 12.83T tokens), Moonshot AI (Kimi K2.6), and MiniMax (M2.7) also performed strongly. Western incumbents Anthropic and Google managed just three spots, looking outmatched both in volume and momentum.
This dramatic shift in traffic share stems from Chinese models' dual breakthroughs in extreme cost-efficiency and long-context processing. DeepSeek V4 Flash's lightweight MoE architecture is a perfect match for autonomous coding agents like Cursor and Claude Code, which demand high-frequency, large-context reads at minimal cost, making it the default choice for developers worldwide in agent scenarios.