Google Launches Gemini 4 Argon, Claims Coding and Security Edge Over Rivals

Google has released Gemini 4 Argon, its most capable frontier model to date, making it available to AI Pro and Ultra subscribers through the Gemini macOS app after a months-long delay that saw the company cede ground to Anthropic and OpenAI. The rollout marks Google’s most aggressive push yet to reclaim leadership in advanced reasoning, software engineering, and defensive cybersecurity — domains where competitors have recently set the pace.

Argon introduces a 1-million-token generation ceiling, a tenfold increase over previous Gemini limits, enabling sustained reasoning chains and end-to-end workflows in a single execution. Google positions the model as purpose-built for software engineering projects, enterprise knowledge work in law and finance, and vulnerability discovery and remediation at scale.

Early benchmarks underscore the specialization. On DeepSWE v1.1, a benchmark measuring performance on real-world software development tasks, Argon posts strong scores that Google says reflect daily internal use — from debugging and large-scale codebase migrations to optimizing quantum computing algorithms and reducing memory footprint in data centers. The company’s own engineers have adopted it as a default tool across those workflows.

The cybersecurity results are more striking. On CWE-bench v1, which tests an AI’s ability to find, validate, and patch software vulnerabilities, Argon outperforms Anthropic’s Claude Opus 5.5 and matches OpenAI’s GPT-6 Astra and xAI’s Grok 4.7 for the top score. Google says the model has already uncovered multiple vulnerabilities across complex codebases spanning 20 programming languages, surpassing the discovery efficiency of its prior 3.8 Flash Cyber model in attack-surface mapping, vulnerability identification, and proof-of-concept generation.

Wiz, the cloud security firm acquired by Google in a $32 billion deal that closed earlier this year, has integrated Argon into its defensive systems — an early signal of how the model may be operationalized across Google’s security portfolio.

Behaviorally, Google says Argon reduces the verbosity and hedging that characterized prior generations, responding more directly and proactively surfacing progress updates during long-running tasks without user prompting. The model was stress-tested for weeks with a specialized group of cybersecurity partners before general availability.

The launch arrives as the frontier model race enters a new phase defined less by chat fluency and more by measurable utility in high-stakes technical domains. With Argon now in subscribers’ hands, the market has its first direct look at whether Google’s delayed entry can translate benchmark leads into sustained developer and enterprise adoption.