MiniMax M2.5 leads Chinese AI surge to 61% OpenRouter share
Cheaper by 17x than Claude Opus 4.6, with nearly matching SWE-Bench scores.
JPMorgan Asset Management strategist Michael Cembalest highlighted in his 2026 Memorial Day commentary that Chinese AI models have taken over global usage rankings on OpenRouter, the largest LLM API aggregation platform serving over 5 million developers. By late February 2026, Chinese models accounted for 61% of token consumption among the top ten models, processing 5.3 trillion out of 8.7 trillion total tokens that week. The top three spots were held by MiniMax M2.5, Kimi K2.5, and GLM-5, with DeepSeek V3.2 also in the top tier. This is not a gradual drift but a structural shift, driven by low-cost open-source models and China's vast domestic market.
MiniMax M2.5, launched in mid-February 2026, saw a 197% surge in token usage after its debut, reaching 4.55 trillion tokens in one month. It scored 80.2% on SWE-Bench Verified (vs. 80.8% for Anthropic's Claude Opus 4.6) at a price of just $0.30 per million tokens versus $5.00—a 17x cost advantage. OpenRouter COO Chris Clark noted that Chinese open-weight models are 'disproportionately heavy in agentic flows run by U.S. developers.' Programming now accounts for over 50% of token usage on OpenRouter, up from 11% in early 2025, and agent-driven workflows generate more than half of all output tokens. By April 2026, Chinese models represented over 45% of all OpenRouter traffic, up from under 2% in late 2024.
- Chinese AI models captured 61% of top-10 model token consumption on OpenRouter by February 2026.
- MiniMax M2.5 achieved 4.55 trillion tokens in one month at 17x lower cost than Claude Opus 4.6.
- Programming and agentic workflows now dominate token usage, driving demand for cost-efficient Chinese models.
Why It Matters
Cost-effective Chinese open-source models are reshaping global AI development and investment strategy.