I tracked which AI models people actually use for a month. The top 5 are all Chinese or open-weight.

Short summary

OpenRouter usage data shows Chinese and open-source models dominating developer production workloads, driven by 50x cost advantages over premium models like Claude Opus. While benchmarks rank US models highest, practitioners choose cheaper alternatives for token-heavy tasks like agents and RAG. The gap between benchmark rankings and real adoption is now significant.

• •Chinese and open-weight models lead real production usage on OpenRouter
• •Price-performance drives choices: DeepSeek costs 50x less than Claude Opus
• •Benchmark rankings don't reflect actual developer adoption patterns

Generated with AI, which can make mistakes.