DeepSeek V3-0324: Stronger Reasoning at Fraction of Western Prices

DeepSeek, a Chinese AI lab, quietly dropped an updated version of their V3 model that caught everyone off guard. On standard benchmarks for math and coding — the kind of problems that separate good models from great ones — it beats GPT-4o. The context window is 128K tokens, enough for large documents or long conversations. Tool use and function calling have been improved, making it more practical for agent-based apps. The kicker? The API costs just $0.27 per million tokens, compared to several dollars for comparable Western models. And it's fully open source under the MIT license, so you can download the weights and run it yourself. The release reignited debates about whether US export controls on advanced chips are slowing China's AI progress — clearly not as much as hoped. It also raised eyebrows about data practices and whether Chinese models can be trusted in sensitive enterprise contexts.