Google announced the Gemini 4 Argon model, an artificial intelligence model designed for long-horizon professional tasks that require multi-step reasoning, such as software engineering, financial research, legal drafting, and cyber defense. The company says the model raises the output limit to one million tokens, compared with 64,000 tokens previously, enabling it to process long workflows and produce extensive outputs within a single task.
Limited rollout before general availability
Google has begun making Argon available to a group of trusted cybersecurity defenders through the Fairwind program. It also says it is participating in the voluntary U.S. government process for access to models before their launch, and intends to gradually expand access after collecting tester feedback and improving safeguards.
Introductory pricing will start at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens. After the introductory period ends, the price will rise to $4 for input and $20 for output per million tokens. Google says broader availability will begin with paid API customers and Google AI Ultra subscribers, followed by developers, enterprises, and consumers.
Performance in programming and enterprise work
Google says Argon achieved a score of 77.9% on the DeepSWE v1.1 benchmark for long-horizon software engineering tasks, and topped the Vals index, which measures economic impact in finance, programming, law, and tax. It also topped Zapier’s AutomationBench benchmark with a score of 51.3%, and achieved 91.7% on LVBench for understanding long videos.
Within Google, the model was used to improve quantum computing algorithms, analyze memory consumption data in data centers, and migrate codebases from C and C++ to Rust. The company says the memory improvements freed more than 300 terabytes after implementation, with total savings estimated at between 500 terabytes and 1 petabyte. In the libgav1 library, Argon replaced approximately 32,000 lines of SIMD code, resulting in a Rust version 2.7 times faster than the previous Rust port while preserving the same video outputs. Google emphasizes that the migrations undergo auditing, testing, and review before being deployed to production.
Special focus on cyber defense
Google trained Argon to autonomously discover, verify, and remediate software vulnerabilities. It says Wiz is using it as part of the Scan for Good initiative to protect public infrastructure, and that the model discovered, in an early trial, a critical vulnerability that exposed sensitive personal data in healthcare software used by hospitals around the world. On the CWE-bench v1 vulnerability remediation benchmark, Argon tied for first place with a score of 68%.
What changes in practice?
Argon represents a shift from using the model for short answers to running extended professional workflows that include analysis, execution, and review. However, the cited results come from Google or its partners and from specific benchmarks, and do not necessarily indicate comparable performance in every production environment. Current availability is also limited, and Google continues to condition public access on safety testing and improvements in resistance to misuse, indirect prompt injection, and misalignment risks.
The company says it is monitoring the model’s behavior and actions, and strengthening the isolation of environments used in training and high-risk evaluations. It will also initially offer Argon without some cyber safeguards to trusted defenders and its internal teams, a decision that increases potential defensive capability but makes user selection and operational oversight an essential part of the rollout’s safety.