DeepSeek V4.1 Flash benchmark results on CyberOffense and NVIDIA Vera Rubin

DeepSeek V4.1 Flash achieved strong performance on two major benchmarks, completing 72.8 percent of hacking objectives on Vals AI's CyberOffense benchmark for under $3 per run and matching OpenAI's GPT-6 Sol at seven times lower cost. Additionally, Artificial Analysis tested the model on NVIDIA's Vera Rubin NVL72 rack, where a single coding agent reached 1,249 tokens per second and parallel agents delivered significantly higher throughput per GPU than previous hardware. Open article →