DeepSeek-V3 and the Qwen2.5 series have achieved performance approaching GPT-4 level on code and structured data generation benchmarks (such as HumanEval and MBPP), while their API call costs are only about 1/10 of GPT-4's.