Google Gemini 4 Argon Explained: From Coding To Cyber Defence, Here’s What Makes It Different

Google announced Gemini 4 Argon, a frontier AI model it says scores 68.9 percent on Vals knowledge-work benchmark versus 63.1 percent for GPT-6 Astra and 65.8 percent for Claude Fable 5.1, and 91.9 percent on Vibe Code Bench versus 90.3 and 89.6 percent for those rivals. CEO Sundar Pichai said on September 30, 2026 that Argon has "frontier safeguards" and is being rolled out to the US government and trusted cyber defenders through Google's Fairwind Program, with general availability to follow. Google-owned Wiz is testing the model under its Scan for Good initiative, where an early demonstration detected a critical flaw in hospital healthcare software that could have exposed sensitive personal data.

Google has announced its next-generation frontier AI model, Gemini 4 Argon, promising major gains in areas like software development, professional workflows and cybersecurity. The new model is a big step forward from its previous artificial intelligence AI systems, and is designed to take on more challenging and complex tasks, the company says. The launch follows reports that Google had been losing ground to competitors like OpenAI and Anthropic, particularly after speculation around a delay to Gemini 3.5 Pro. Google CEO Sundar Pichai said internal teams are already using Gemini 4 Argon for things like coding and quantum computing. It’s also the first time Google has used the name for an AI model, called Argon. "Importantly Argon has frontier safeguards and we are rolling it out responsibly - it’s with the US govt and going to a set of trusted cyber defenders through our Fairwind Program today,” Sundar Pichai wrote on X. “We’re going to make it available as soon as we can and as safely as we can. So hold tight, lots more coming, and you’re going to see us iterating rapidly." Importantly Argon has frontier safeguards and we are rolling it out responsibly - it’s with the US gov’t and going to a set of trusted cyber defenders through our Fairwind Program today. We’re going to make it available as soon as we can and as safely as we can. So hold tight, lots more coming, and you’re going to see us iterating rapidly. https://t.co/8MiXIfXIKr — Sundar Pichai @sundarpichai September 30, 2026 https://x.com/sundarpichai/status/2105387954474746264?ref src=twsrc%5Etfw How Capable Is Gemini 4 Argon? Google claims that Gemini 4 Argon is a pioneer model that deals with advanced reasoning and multi-step workflow capabilities. It includes output capacity up to 1 million tokens, which helps the model to sustain longer reasoning trails and produce hundreds of thousands of tokens while performing difficult tasks. Based on Google benchmarking results, Gemini 4 Argon can compete with and even outperform leading models such as GPT-6 Astra and Claude Fable 5.1. For instance, its score on Vals knowledge-work benchmark is 68.9 percent. At the same time, GPT-6 Astra gained 63.1 percent, and Claude Fable 5.1 – 65.8 percent. The model performed well during coding evaluations too. The score of Gemini 4 Argon on Vibe Code Bench was 91.9 percent, whereas the scores of Claude Fable 5.1 and GPT-6 Astra were 90.3 and 89.6 percent correspondingly. However, the model does not perform better than its competitors in all benchmarks. For example, Argon performed worse than competing models in Terminal-bench 4.0. Gemini 4 Argon’s Cybersecurity Capabilities There are further enhancements in cybersecurity, which Google claims that Gemini 4 Argon has achieved. The company claims that the model is trained to have robust cybersecurity defence capability and is capable of autonomously detecting and fixing important vulnerabilities in the software. Cybersecurity stocks are trading lower after Google unveiled Gemini 4 Argon, which is highly capable at cyber defense & will be offered to trusted defenders without cyber guardrails OpenAI execs have also said they’re actively building cybersecurity products around Astra https://t.co/dTget3I2r3 — Wall St Engine @wallstengine October 1, 2026 https://x.com/wallstengine/status/2105633408626135445?ref src=twsrc%5Etfw The cybersecurity firm Wiz, owned by Google, is conducting tests on Argon under the Scan for Good initiative of the firm. In one of the early demonstrations, the model detected a critical flaw in the healthcare software that could allow the exposure of sensitive personal data in the software used in hospitals across the globe. Gemini 4 Argon Gets a Limited Release Google is using a restricted release strategy amid fears of growing risks related to AI safety and cybersecurity. The company is giving Argon access to chosen cyber defenders under its Fairwind project and the US government. Google’s strategy resembles that adopted by other AI companies for the access to their models such as OpenAI's Daybreak and Anthropic’s Project Glasswing. Google vient de dévoiler Gemini 4 Argon, son IA la plus avancée ... et vous n'avez pas le droit de l'utiliser moi non plus . Elle sort d'abord chez un premier groupe d'experts en cybersécurité, et le gouvernement américain l'examine avant tout le monde. Eux l'ont même reçue sans ses garde-fous cyber. Pourquoi autant de précautions ? Parce qu'elle trouve des failles que les autres ne voient pas. La société Wiz s'en est servie sur un logiciel utilisé par des hôpitaux du monde entier : elle y a trouvé une faille critique qui exposait des données personnelles sensibles, et que les IA les plus avancées avant elle avaient ratée. Chez Google, des milliers d'employés l'utilisent déjà : → Les chercheurs en informatique quantique lui ont confié un bout de programme à alléger. En quelques minutes, elle a fait 40 % mieux que la solution de référence publiée → Elle a permis de libérer plus de 300 To de mémoire dans les data centers de Google la RAM de près de 20 000 PC portables → Elle peut écrire jusqu'à 1 million de tokens en une seule réponse, contre 64 000 avant Bon d'après Bloomberg, des employés de Google la trouvent moins bonne sur leurs vraies tâches de code que sur les benchmarks. Google dément. MAIS sur le classement indépendant d'Artificial Analysis, elle fait jeu égal avec GPT-6 Astra Claude Opus 5.5 reste devant . Et quand elle ne connaît pas la réponse, elle invente dans 15 % des cas, contre 51 % pour GPT-6 Astra. Bientôt la fin des mèmes sur le retard de Google ? — VISION IA @vision ia October 1, 2026 https://x.com/vision ia/status/2105638485445030208?ref src=twsrc%5Etfw More access will come with paid API users and Google AI Ultra subscribers once the model has been tested by trusted testers and cybersecurity experts. Google has fixed an introductory API rate of $2 or around INR 190 per million input tokens and $10 or around INR 950 per million output tokens. The cached input tokens will be charged at 95 per cent discount. Post introductory period, the prices will go up to $4 or INR 380 per million input tokens and $20 or approximately INR 1,900 per million output tokens. Also See: Gemini 4 Argon Gets Limited Rollout: Google Reveals New Features, Pricing and Usage Limits https://in.mashable.com/tech/114498/gemini-4-argon-gets-limited-rollout-google-reveals-new-features-pricing-and-usage-limits