The announcement, published by Google on September 30, introduces the first model in its Gemini 4 generation. Google says Argon can sustain reasoning across long tasks, including security vulnerability discovery and patching, legal and financial research, and complex coding. The company has not set a public launch date. It plans to begin broader access with paid API customers and Google AI Ultra subscribers after the early testing phase.
The phased launch puts security testing at the center of the product release. Google says it is participating in the U.S. government’s voluntary pre-release model access process and will use feedback from early testers to refine protections before making Argon more widely available. CNBC reported the initial group will be trusted cybersecurity partners, with safeguards still being developed to address misuse and prompt injection.
A much larger output window for long tasks
Google is raising Argon’s maximum output length to one million tokens, up from 64,000 in the previous generation, according to its announcement. That capacity is intended for work where a model needs to carry a lengthy reasoning process through a single response. It could support analysis spanning large codebases or lengthy business material, though the context limit alone does not show how accurately the model completes such tasks.
The model also handles visual information, Google says. It can analyze charts, retrieve details from long videos and use information across multiple documents. Those features widen the kinds of material that can be included in a single workflow. Google describes Argon as useful for coding, research and enterprise processes, but those are company claims at launch. Independent users cannot yet test the system through a public API or consumer product.
Google’s benchmark claims and internal use
Google reports a score of 77.9% on DeepSWE v1.1, a test of long-running software engineering work, and 51.3% on Zapier’s AutomationBench, which evaluates business task execution. It also says Argon leads the Vals Index, a measure of model performance across finance, coding, legal and tax tasks. These numbers come from evaluations chosen and reported by Google. They should be read as benchmark results, not proof that Argon will outperform competing systems in every organization’s software or office workflows.
The company says Argon has already been used internally by Google employees for debugging, algorithm design and code migration. One reported project involved converting C and C++ software to Rust, including work on the Fuchsia operating system’s Zircon kernel. Google says those rewrites are subject to automated and manual audits, emulation testing and review before deployment. That qualification matters for companies considering AI-assisted changes to production code: generated code still needs security and reliability checks.
Google also describes using agents powered by Argon to analyze data-center memory telemetry and apply optimizations. It says the work freed more than 300 tebibytes of memory after rollout, with estimated total savings of 500 tebibytes to one pebibyte. That is a company-reported internal result rather than a third-party measurement, and Google has not said how the result compares with alternative optimization methods.
Cybersecurity access comes with a trade-off
Argon is designed to find, validate and patch software vulnerabilities, according to Google. For its early trusted defenders and internal teams, the company says it will provide the model without cyber guardrails so they can use its full defensive capabilities. Broadening access to a system with vulnerability discovery abilities could help security teams inspect software faster, but the same capabilities could be misused. Google says its remaining work includes preventing harmful use and improving resilience against prompt injection, where hostile instructions in content can redirect a model’s actions.
That makes access controls and independent evaluation important parts of the rollout, especially for businesses that may eventually connect an advanced model to code repositories or internal tools. The company says it will continue to test the system and improve its safeguards as it expands access. CNBC reported that the model is also being assessed with the U.S. government before a public launch.
Google lists introductory API prices of $2 per million input tokens and $10 per million output tokens, with cached input priced 95% below the standard input rate. The company says those introductory rates will later rise to $4 and $20 per million input and output tokens. The final availability date, the access conditions for customers and how the model performs on independent evaluations remain open questions.
Sooraj Raj is a programmer, blogger and professional developer with an interest in technology research and publishing. He writes about technology developments and shares information in a clear, straightforward way.
