I ran an eval to choose between DeepSeek V4 Flash and GPT-6 Luna for one of my projects, and I learned something new. On paper, DeepSeek looks cheaper. The input price is about the same: $0.10 per million tokens for Luna and $0.09 to $0.14 for DeepSeek. Luna's output price is up to 2.8 times higher: $0.50 per million tokens against $0.18 to $0.28 for DeepSeek. In practice, Luna cost less per task, because it is much more token efficient. DeepSeek spent 703 hidden reasoning tokens per call and Luna spent 120. Per call: $0.000385 for DeepSeek and $0.000283 for Luna. Median response time: 24s for DeepSeek and 3.5s for Luna. By the way, running an eval is a single prompt now. This was mine: "Compare DeepSeek V4 Flash and GPT-6 Luna on the top tools in this project, based on usage. Run an eval suite with a judge on my Claude Code subscription, and keep it under $1."