Open-weight DeepSeek model outperforms Google's gated Gemini in bug hunting

The security gate dilemma

Google recently introduced Gemini 4 Argon, its most capable model to date. However, the company took an unusual step by initially refusing to release it to the general public. Instead, they directed the model only to a limited circle of verified security professionals through a program called Fairwind.

The company expressed concerns that Argon was exceptionally good at finding software vulnerabilities. They argued that this specific skill could be used by bad actors to break into sensitive systems like hospitals or government networks. Google DeepMind's chief AI architect stated that "safely releasing frontier capabilities at this level requires a phased approach."

This cautious approach was intended to prevent the model from being used for cyberattacks. By keeping the model behind a "velvet-rope" program, Google aimed to ensure that only authorized defenders could utilize its advanced reasoning. However, recent independent testing suggests that this gate may not be as effective as intended.

Open-weight performance

While Google restricted its model, a rival company released DeepSeek V4 Pro as an open-weight model. This means anyone can download and run the model on their own hardware without seeking permission from a vendor. Despite being freely available, it has shown significant capabilities in identifying exploitable flaws in real code.

A cybersecurity firm called Aikido tested ten different AI models against 32 freshly disclosed CVEs. CVEs are common vulnerabilities and exposures that are publicly known. The goal was to test genuine reasoning rather than the model simply recalling answers it had already seen during its training process.

DeepSeek V4 Pro topped every public closed model in the test. By pooling three runs, it rediscovered 28 of the 32 vulnerabilities, resulting in an 87.5% hit rate. This outperformed closed models from Anthropic and Alibaba, which both reached a lower score of 81.3%.

Cost and accessibility

The performance gap is accompanied by a significant difference in pricing. Aikido reported that three runs of the DeepSeek model cost roughly $295 total. In contrast, a single pass from some of its competitors, such as Grok 4.6 or OpenAI's Sol, cost between $450 and $590.

DeepSeek also demonstrated high performance on other coding benchmarks. It scored 80.6% on SWE-bench Verified and achieved the highest Codeforces rating of any model tested in the study. Furthermore, it is priced at roughly 46 times less than Claude Opus 4.7 when looking at a blended token rate.

Other models are also showing strong security performance in the open-weight space. Models from Kimi and Zhipu's GLM-5.3 showed consistent results in vulnerability finding. These are not just fringe experiments; they are tools that security teams can already run today without any vendor gate.

The shifting timeline

The duration of Google's restriction on Gemini 4 Argon was notably short. Although it was initially held back for a government pre-release review, it reached general availability on October 13. This was only two weeks after the restricted rollout began for cybersecurity defenders.

Google eventually made the model available on the paid tier of its API and to subscribers of its Gemini app and Search. This move suggests that the company may not have viewed the risk as severe enough to maintain a long-term hold. The time between restricting the model to security experts and making it available to anyone with a credit card was only about fourteen days.

While Gemini 4 Argon remains a strong model, the data shows it is not uniquely dangerous compared to its rivals. It leads or ties in 13 of 18 categories on Google's own benchmarks. However, the fact that an open-weight model can already perform the same tasks at a lower cost complicates the argument for restricted access.

What comes next

The success of DeepSeek V4 Pro highlights a growing trend where open-weight models are reaching parity with closed systems. This puts pressure on companies like Google and Anthropic to justify why they keep their most capable models behind paid APIs or vetted programs. If open models can do the work, the "gatekeeper" model of AI safety faces a difficult path forward.

Security teams are already beginning to integrate these open models into their workflows. Because these models can be run on private hardware, they offer a level of privacy and control that gated APIs cannot provide. This could lead to a shift where the most sensitive security work is done entirely off-platform.

Moving forward, the industry will likely see more benchmarks focusing on vulnerability reasoning. As models become better at finding bugs, the focus will shift from whether the models *can* do it to how the public can be protected from misuse. The two-week window for Gemini 4 Argon may be a sign of how quickly these capabilities are becoming mainstream.