漏洞挖掘实测：GPT-6 Sol 与 5.6 Sol 行为差异

BradMoon @xy9301原推发布于 2026.09.23

在 X 上打开原推

When hunting for vulnerabilities, GPT-6 Sol tends to focus on validating and wrapping up a single provable finding, while GPT-5.6 Sol is more willing to keep asking, “What else could go wrong in this same code?” In this round of testing, that made 5.6 Sol more likely to uncover bugs hiding in edge cases. But the sample was tiny, so this is only a tendency observed in this round. Cost down 60%, time down 30%, input down 15%, output down 30%, the same direction as the DeepSeek v4.1 update. Essentially, astra = sol, sol = terra. OpenAI makes you think you’re using Sol, but you’re not.