The original title is: "Production data: Chinese vs US LLM API pricing, latency, and benchmark comparison"

Short summary

An infrastructure architect shares 18 months of production data comparing US (OpenAI, Anthropic, Google) vs Chinese (DeepSeek, Qwen, GLM, Kimi) LLM APIs. Chinese models like DeepSeek V4 Flash are 20-60× cheaper than US equivalents with comparable latency (p99 ~1.1s) and uptime (99.95%+). Benchmark gaps of 1-3 points on MMLU and HumanEval rarely justify the cost differential, making Chinese models viable for 70% of production traffic.

• •Chinese LLM APIs are 20-60× cheaper than US equivalents with comparable latency and uptime
• •DeepSeek V4 Flash handles 70% of author's production traffic at $0.25/M output tokens vs $10-15/M for US models
• •Benchmark quality gaps of 1-3 points rarely justify the cost premium for most workloads

Generated with AI, which can make mistakes.