What does DeepSeek still do better than every other model for you?

Forget benchmarks. What's the one real-world task where you tried GPT, Claude, Gemini, Qwen etc. and still came back to DeepSeek? Bonus points if you have a prompt that makes the difference obvious.