DeepSeek-V4-Pro Keeps Getting Stronger: Three Runs Recover 87.5% of Vulnerabilities, Surpassing Claude Opus 5 and Qwen3.8 Max

According to monitoring by Beating, security company Aikido tested 32 real-world vulnerabilities using 8 cutting-edge models. DeepSeek-V4-Pro-0813 could only identify an average of 58.3% in a single run, but after running three consecutive times and merging the results, it directly identified 28 vulnerabilities, covering 87.5%, jumping to the top spot. Claude Opus 5 and Qwen3.8 Max only achieved 81.3%. Even more striking is the cost. On average, DeepSeek spends only $7.34 to identify one vulnerability, while Opus 5 costs $66.97—DeepSeek's cost is approximately 1/9 of Opus 5's. DeepSeek's performance is also unusual: single runs are unstable, but running multiple times yields the best results. Each run identifies different vulnerabilities, and after three runs, the combined coverage surpasses all competitors. The trade-off is a high false positive rate. Among the issues reported by DeepSeek, only 65.6% were ultimately confirmed as real vulnerabilities, the lowest among the 8 models tested. [Original Link]