Google Unveils Gemini 4 Argon, Restricts Initial Access to Trusted Cyber Defenders

Google DeepMind said Argon is the first model in the Gemini 4 generation and can generate up to 1 million output tokens in a single response, up from 64,000 on earlier Gemini models. It is designed for long-horizon coding tasks, legal and financial analysis, and cyber defense. Google said it is taking a phased approach, starting with trusted cybersecurity partners, while participating in the U.S. government’s voluntary process for pre-release model access. The company said it will collect feedback from early testers and strengthen guardrails before a wider release.

The launch came a day after Google CEO Sundar Pichai signed a voluntary safety accord with President Donald Trump following a White House meeting with major technology executives on AI safety, CNBC reported. Google said it is working on pre-release safety evaluation with the U.S. government and plans safeguards in four areas: misuse defenses for cyber and CBRN risks, resistance to indirect prompt injection, misalignment monitoring of chain-of-thought and actions, and sealed sandboxes for high-risk training and evaluations. The model leads Gray Swan’s indirect prompt injection benchmark, according to Google.

Google said Argon leads outright on 12 of 18 benchmarks and ties for first on one, compared with OpenAI’s GPT-6 Astra, Anthropic’s Claude Opus 5.5 and Claude Fable 5.1. The company reported 77.9% on DeepSWE v1.1 for long-horizon software engineering, 68.9% on the Vals Index for economic impact across finance, coding, legal and tax, 51.3% on AutomationBench, 19.6% on the Harvey Legal Agent Benchmark and 91.7% on LVBench for long video understanding. It trails on some tests: 55.0% on FrontierSWE v2 against GPT-6 Astra’s 65.5%, 57.4% on Terminal-Bench 4.0 against Claude Opus 5.5’s 66.4%, and 69.2% on OSWorld-2.0 against GPT-6 Astra’s 72.6%. On CWE-bench v1, which tests vulnerability remediation, Google said Argon ties for first at 68%. CNBC reported that Argon also ties with OpenAI’s GPT-6 Astra and Grok 4.7 on cybersecurity evaluation benchmarks.

Wiz is already using Argon through its Scan for Good initiative, and the model found a critical vulnerability in healthcare software used by hospitals worldwide that previous frontier models had missed, according to MarkTechPost. Google said trusted defenders and internal teams receive Argon without cyber guardrails.

Inside Google, thousands of employees already use Argon, the company said. Argon agents applied memory optimizations across data centers, freeing more than 300 TiB of memory, with 500 TiB to 1 PiB projected. The company also said agents replaced 32,000 lines of SIMD code in the libgav1 Rust port, making the decoder run 2.7 times faster with identical output, and are migrating C/C++ codebases to Rust, including more than 800,000 lines in the Fuchsia OS Zircon kernel. Quantum computing researchers have used the model, and Argon beat a published quantum algorithm baseline by 40% in minutes, Google said. CNBC described the memory savings as hundreds of terabytes.

Pricing is public. Argon launches at introductory rates of $2 per 1 million input tokens and $10 per 1 million output tokens, with cached input tokens discounted 95% to $0.10 per 1 million. After the introductory period, pricing moves to $4 per 1 million input tokens and $20 per 1 million output tokens. Google did not disclose Argon’s input context window. The model will initially be available through the Fairwind Program, then expand to Google AI Ultra subscribers, Android Authority reported. Google has not given an exact release date beyond “soon.”

The release puts Google back at the frontier nearly a year after Gemini 3, following a year in which the company focused on faster, lower-cost Flash models. It also arrives shortly after OpenAI’s DevDay conference, where OpenAI launched its Dots AI agent and GPT-6.1 Sol model, and after Koray Kavukcuoglu became the new head of DeepMind in August, The Verge reported. Google’s claims are based largely on its own published comparisons and benchmark results; independent hands-on testing of Argon is not yet available.

Editor's Summary Google announced Gemini 4 Argon, a frontier AI model for coding, legal and financial work, and cyber defense, but limited initial access to trusted cyber defenders. Google reported leading scores on several benchmarks, disclosed API pricing and described internal uses including memory savings and code migration. Broader access will expand to Google AI Ultra subscribers and eventually other users, though no exact date was given.