Google Gemini 4 Argon Unveiled With 1 Million Token Output Limit and Advanced Cybersecurity Capabilities

Google DeepMind has announced Gemini 4 Argon, a new frontier AI model designed to handle complex, long-running tasks across software engineering, finance, legal work, visual analysis and cybersecurity.

Unlike a conventional public launch, Google is initially making Argon available to a limited group of trusted cyber defenders through its Fairwind Program. The company plans to expand access gradually, beginning with paid API customers and Google AI Ultra subscribers. A specific public release date has not yet been announced.

Gemini 4 Argon Built for Long and Complex Tasks

One of Argon’s key features is its expanded output capacity. Google has increased the maximum output limit to 1 million tokens, compared with 64,000 tokens for earlier Gemini models.

The larger limit is intended to help the model maintain reasoning over lengthy, multi-step tasks, including large software projects, extended research and complex enterprise workflows.

Google said its own engineers are already using Argon for debugging, algorithm development and large-scale code migrations. The company said Argon agents are helping migrate C and C++ codebases to Rust, including work involving more than 800,000 lines of code in the Zircon kernel of Fuchsia.

Google also reported that Argon agents identified memory-efficiency improvements across its data centres. The company said more than 300 TiB of memory could be freed once the changes are rolled out, with estimated total savings of between 500 TiB and 1 PiB.

Coding, Finance and Legal Performance

Google reported a score of 77.9 per cent for Gemini 4 Argon on DeepSWE v1.1, a benchmark focused on long-horizon software engineering tasks. The company also reported strong results in finance, legal research, business automation and long-video understanding.

Google said Argon leads its Vals Index, which evaluates work across finance, coding, legal and tax tasks. It also reported a top score of 51.3 per cent on AutomationBench and 91.7 per cent on LVBench, which measures long-video understanding.

However, these results are based largely on Google’s own evaluations. Early independent coverage has noted that Argon does not lead every benchmark, with some competing models performing better on certain terminal-based and software-engineering tasks.

Cybersecurity Access Restricted at Launch

Cybersecurity is a major focus of Gemini 4 Argon. Google says the model can autonomously identify, validate and patch critical software vulnerabilities.

Trusted cyber defenders and Google’s internal teams will initially be able to use Argon with fewer cyber-related restrictions, allowing them to test its capabilities in real-world defensive environments.

Wiz is among the early users through its Scan for Good initiative. Google said Argon identified a critical vulnerability involving the exposure of sensitive personal information in healthcare software used by hospitals worldwide. The company said previous frontier models had missed the issue.

Google is also adding safeguards against misuse, including protections against harmful cyber and biological requests, prompt-injection attacks and behaviour that goes beyond a user’s intended instructions. The company said monitoring systems can track the model’s reasoning and actions and stop execution when necessary.

Gemini 4 Argon Price and Availability

Google has announced an introductory API price of $2 per million input tokens and $10 per million output tokens. Cached input tokens will receive a 95 per cent discount from the input-token price. After the introductory period, Google says pricing will rise to $4 per million input tokens and $20 per million output tokens.

For now, Gemini 4 Argon remains unavailable for general public use. Google says broader access will begin with paid API customers and Google AI Ultra subscribers before expanding to developers, enterprises and consumers. No firm public launch date has been announced.

Disclaimer

This article is based on information released by Google DeepMind and early reporting available at the time of publication. Benchmark results cited by Google may differ from results obtained through independent testing, and the model’s availability, pricing and capabilities may change as the rollout progresses. Readers should refer to Google’s official announcements for the latest information.