DeepSeek V4-Flash hands-on: Costs cut by 40%, but did performance suffer?

A friend recommended DeepSeek V4-Flash, saying it costs only 40% of GPT-5.6 Luna but performs on par in benchmarks. My first thought was: another "cost-performance" boast? As an investor, I hate this kind of sales pitch. But since it's my line of work, and I've been watching exit paths for AI infrastructure lately, it's necessary to give it a hands-on test.