The cyber capability gap between open and closed AI models keeps shrinking

What the benchmark shows

The Vals AI test measured how far models could get toward set attack objectives in simulated corporate networks. OpenAI's GPT-6 Sol Daybreak hit 80.8 percent of those objectives.

The more striking result came on the open-weight side: DeepSeek V4.1 Flash reached 72.8 percent, coming close to the closed model at roughly a tenth of the cost.

Why it matters

Once open-weight models are released, there is no way to restrict who uses them. As the capability gap narrows, the time defenders have to prepare shrinks with it.

Analysts say the window for defenders is closing fast. In other words, as attack capability rises, organizations have less time to harden their networks.

Limits of the numbers

The results are not a one-to-one picture of real-world attacks; they come from a controlled simulation. Still, they are meaningful for showing that the distance between open and closed models is closing despite the cost difference.