--- title: "Google Announces Gemini 4 Argon, Its First Frontier Model in Months" description: "Only vetted cybersecurity partners can use the model at first, though Google claims benchmark wins over OpenAI and Anthropic." author: "rews desk" published: 2026-10-01T14:39:37Z modified: 2026-10-02T03:18:04Z url: https://rews.cc/a/google-announces-its-first-gemini-4-model-with-benchmark-win-fe3d20 language: en tags: ["google", "ai", "gemini", "cybersecurity", "openai", "tech"] publisher: "Rews (https://rews.cc)" --- # Google Announces Gemini 4 Argon, Its First Frontier Model in Months *Only vetted cybersecurity partners can use the model at first, though Google claims benchmark wins over OpenAI and Anthropic.* By rews desk · October 1, 2026 · https://rews.cc/a/google-announces-its-first-gemini-4-model-with-benchmark-win-fe3d20 ## In brief - Google announced Gemini 4 Argon, its first new frontier AI model in months - Initial access is limited to vetted cybersecurity partners in Google’s Fairwind Program - Google says Argon leads or ties 13 of 18 benchmarks against GPT-6 Astra and Claude Opus 5.5 - Inside Google, Argon has found more than 300 tebibytes of data-center memory savings - Output limit rises to 1 million tokens, with introductory API pricing of $2 and $10 per million tokens Google on Wednesday announced Gemini 4 Argon, its first new frontier artificial intelligence model in months, and immediately restricted it to a small group of vetted security partners while the company works through a U.S. government review process for advanced models. Access begins through the Fairwind Program, which lets approved governments and cyber authorities test Google models early, and reaches what the company calls “trusted cyber defenders,” who can run the model without its usual cyber safety restrictions. Koray Kavukcuoglu, the head of Google DeepMind, said Google was “actively engaged” in the government’s voluntary pre-release access process while it gradually widens availability. Broader release, starting with paying API customers and Google AI Ultra subscribers, is planned “as soon as possible,” the company said. The model is Google’s bid to rejoin a frontier race it has been losing. A flagship successor, Gemini 3.5 Pro, was promised for June, then postponed repeatedly over performance problems, and never shipped. Google has since released only smaller Flash models while OpenAI and Anthropic set the pace in coding and cybersecurity. Google published results for 18 benchmarks comparing Argon with OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5. Argon leads outright in 12 categories and ties for first in one. On DeepSWE v1.1, a test of long software-engineering jobs, it scored 77.9 percent, against 74.2 percent for Opus 5.5 and 74.1 percent for Astra. On Harvey’s legal-agent benchmark it scored 19.6 percent to Astra’s 5.4 percent. But Astra beat Argon by 10.5 points on both FrontierSWE v2 and Terminal-Bench Science, and Opus 5.5 led Terminal-bench 4.0 by nine points. Argon also scored 68 percent on CWE-bench, a vulnerability-remediation test, which Google’s materials show as a tie with Astra for first place. InfoWorld reported that the materials also listed Opus 5.5 and xAI’s Grok 4.7 at the same score. Inside Google, thousands of employees already use the model, the company said. Argon analyzed fleet-wide telemetry to find more than 300 tebibytes of memory savings across Google’s data centers, with another 500 tebibytes to 1 pebibyte estimated to follow, and it is converting C and C++ code to Rust, including more than 800,000 lines in the Fuchsia operating system’s Zircon kernel. Quantum-computing researchers used it to beat a published efficiency baseline by 40 percent. In security work, the firm Wiz is already scanning public infrastructure with Argon under a program that fixes what it finds free, and Google said the model caught a serious flaw in hospital software used worldwide, one that exposed personal data and that earlier frontier models had missed. “We really believe that a model of this caliber and this level of frontier performance is meaningfully important for defenders,” Tulsee Doshi, the Gemini product lead, told CNBC. Before any wider release, Google says, it is blocking misuse for cyberattacks while permitting defensive research, hardening Argon against indirect prompt injection, sealing the sandboxed environments used for high-risk testing, and watching the model’s chain of thought in real time, stopping it when it strays from a user’s instructions. What that monitoring catches will not be fed back into training, the company said, so Argon does not learn to reason its way past its watchers. Argon also raises Google’s output limit to 1 million tokens, from 64,000 in earlier Gemini models, letting it sustain a long piece of work in a single pass. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached input. Google has not said when the rates will rise to $4 and $20, the prices Anthropic charges for Opus 5.5, and less than half of what OpenAI charges for GPT-6 Astra, at $10 and $50. The announcement came a day after OpenAI’s annual developer conference, where it introduced an agent called Dots and a model called GPT-6.1 Sol and confirmed it would not release a planned GPT-6.1 Astra model because of safety concerns. Analysts stopped short of crowning Argon. “I would call it Google’s frontier-class model,” said Stephanie Walter of HyperFRAME Research, “but its mixed benchmark results and limited availability make it too early to declare it the new market leader.” ## Sources - [Google announces Gemini 4 Argon AI model, but you can't use it yet](https://arstechnica.com/google/2026/09/google-announces-gemini-4-argon-ai-model-but-you-cant-use-it-yet/) — Ars Technica - [Google launches Gemini 4 Argon to reclaim the AI frontier](https://www.businessinsider.com/googles-gemini-4-argon-frontier-engineering-knowledge-work-cyber-defense-2026-9) — Business Insider - [Google announces Gemini 4 and says it’s so capable that only ‘trusted cyber defenders’ can have it right now](https://theverge.com/tech/1002980/google-gemini-4-argon) — theverge.com - [Google unveils Gemini 4 Argon, retaking benchmark lead over OpenAI and Anthropic — but in limited release](https://venturebeat.com/technology/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic-but-in-limited-release) — venturebeat.com - [Google Debuts Gemini 4 Argon AI Model with Restricted Cyber Defender Launch](https://techstrong.ai/agentic-ai/google-debuts-gemini-4-argon-ai-model-with-restricted-cyber-defender-launch/) — techstrong.ai - [Inside Gemini 4 Argon, the model Google is testing on its own infrastructure first](https://securityaffairs.com/200187/uncategorized/inside-gemini-4-argon-the-model-google-is-testing-on-its-own-infrastructure-first.html) — securityaffairs.com - [Google makes Gemini 4 AI model available to a trusted few](https://infoworld.com/article/4229615/google-makes-gemini-4-ai-model-available-to-a-trusted-few.html) — infoworld.com