DeepSeek-V4 Flash vs GPT-5.6 Luna: cost and coding on DeepSWE

On 900 DeepSWE rollouts, GPT-5.6 Luna leads pass@1 at 67.2% vs DeepSeek-V4 Flash's 53.3%, but DeepSeek costs $0.10 per task vs Luna's $0.61. A DeepSeek-first cascade solves 78.9% of tasks at $0.385 each, beating Luna alone and 37% cheaper.