DeepSeek Nears OpenAI's Cyber Model on Vals Test at About $3 a Run

A new leaderboard from evaluation firm Vals AI, reported on November 8 by FourWeekMBA, places an open-weight DeepSeek model close to OpenAI's cyber model on a network-attack benchmark, at roughly $3 per run.

The test, called CyberOffense, does not ask models to recite vulnerability knowledge. It drops AI agents into two simulated company networks that are isolated from real systems and lets them work without human help, hunting for entry points and moving through attack steps on their own. Objectives are fixed in advance, and a result only counts when an agent reaches a hidden marker that proves it got there, not when the model claims success.

According to the leaderboard as relayed by FourWeekMBA, OpenAI's cyber model still finishes first overall and DeepSeek follows, with the total gap compressed to single-digit percentage points. On the two simulated networks the two systems trade places, each leading on one of them.

The cost line is what drew attention. The DeepSeek model runs for under $3 per attempt, while the OpenAI model, which the report says is offered only through the company's cyber program, costs several times more per run. For red teams that need to iterate repeatedly and score many runs, that price difference decides who can afford to use this class of capability at all.

Vals describes the DeepSeek system as an open-weight model that anyone can download. That detail pushes the result past a simple ranking. The gap that once separated a restricted frontier cyber model from a freely downloadable one now looks much narrower.

The caveats are just as important. Vals compares models at equal compute, counting only objectives completed inside a fixed token budget, so cost and score have to be read together. A score on simulated ranges is also not the same as the ability to break into a real network, and one leaderboard cannot capture how stable a model is across environments.

For defenders, Vals frames the finding simply: the same models that find a way into a network also find the weaknesses that let them in. The practical use of such a benchmark may be less about ranking vendors and more about letting companies run automated red-team exercises against their own systems.

Three things to watch from here: whether the leaderboard order shifts as new models are added, whether DeepSeek holds its near-parity across more network targets, and whether OpenAI adjusts how it gates access to its cyber model in response to public comparisons like this one.