Gemini 4 Is Here, and Google’s Flagship Tops All Other AI Models on Cybersecurity

Google unveiled Gemini 4 Argon, positioning it as the frontier model for tasks including coding, office work, and cyber defense. Argon scored 77.9% on the DeepSWE v1.1 benchmark, positioning it ahead of competitors like Claude Opus 5.5 (74.2%) and GPT-6 Astra (74.1%). In a prompt injection test by Gray Swan's Indirect Prompt Injection benchmark, Argon recorded a 0.7% attack success rate, better than Claude Opus 5.5 and Claude Fable 5.1, both of which scored 1.0%. The model is made available to vetted cyber defenders through the Fairwind Program before being released to paid API customers and Google AI Ultra subscribers. Argon can generate up to 1 million tokens in a single reply, an increase from the previous 64,000.

The narrative frames the release of Gemini 4 Argon not purely as an advancement in capability but as a strategic deployment of security-focused testing, positioning the model as a tool to understand adversarial thinking for defensive purposes. The focus on prompt injection scores reveals a subtle hierarchy where Argon demonstrates superior resistance against specific, targeted adversarial attacks compared to its immediate competitors. This suggests that optimizing for security resilience, rather than raw performance metrics like DeepSWE, is emerging as a significant differentiator in advanced AI development. Furthermore, the decision to release the model through a phased rollout—prioritizing vetted defenders before broader public access—establishes a paradigm where adversarial modeling informs deployment. The context of the surrounding events, including the subsequent reports on other large models, suggests an underlying tension regarding the pace and control of AI advancement by major labs. The implication is that defensive utility may be prioritized over immediate competitive advantage in the general release phase.