Freelancer Cuts AI API Costs 40x by Switching Routine Tasks to Chinese Models

A freelancer discovered they had wasted around $600 of a client's budget by using expensive models like GPT-4o for simple tasks such as CSV cleaning and product description rewrites. To find a cheaper alternative, they spent two weekends routing real client work through Chinese AI models — DeepSeek, Qwen, Kimi, and GLM — via a unified API endpoint, tracking every cost and quality metric. DeepSeek V4 Flash, priced at $0.25 per million output tokens, delivered output quality comparable to GPT-4o at roughly 40 times lower cost on identical tasks. The freelancer found the model fast enough for real-time client workflows and competitive on coding benchmarks, though it lacks native vision support and is not the top choice for China-targeted marketing copy. Other models in the DeepSeek lineup — including V3.2, V4 Pro, and R1 — were used selectively depending on task complexity and quality requirements.