Gemini 4 Argon

TL;DR

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get tech for your team delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Google announced Gemini 4 Argon on Sept. 30, saying it is rolling out first to trusted cyber defenders through its Fairwind Program. The company reports strong results in coding, enterprise tasks and defensive cybersecurity, but those results are company claims; broader availability and independent validation remain pending.

Google announced Gemini 4 Argon on Sept. 30, describing it as a frontier model for extended software engineering, enterprise work and defensive cybersecurity. The company said access is beginning with trusted cyber defenders in its Fairwind Program, with broader release planned after feedback and further work on safeguards.

Google says Argon is built to sustain reasoning over long, multi-step workflows. It has expanded the model’s output limit to 1 million tokens, compared with 64,000 for the previous limit cited in the announcement. Google says this gives the model room to work through larger tasks in a single run. The announcement does not provide independent evidence about how reliably it performs those tasks.

The company reports that Argon is already used by thousands of Google employees for coding, research and writing. Google describes internal projects involving memory optimization across its data centers and migrations of large C and C++ codebases to Rust. It says one Rust video decoder project replaced 32,000 lines of SIMD code and ran 2.7 times faster than an existing Rust port while producing identical video output. Google says these systems undergo automated and manual audits, emulation tests and review before production use.

Google also cites results from outside evaluations: 77.9% on DeepSWE v1.1 for long-horizon software engineering, 51.3% on Zapier’s AutomationBench, and 91.7% on LVBench for long-video understanding. The company says Argon leads on the Vals Index and names finance and legal benchmarks as further evidence. These are reported benchmark results; the announcement does not detail independent replication or establish that benchmark performance will match results in everyday use. Google lists planned introductory prices of $2 per million input tokens and $10 per million output tokens, with cached inputs priced 95% below the standard input rate.

A Restricted Start for Cyber Defense

The staged release puts cybersecurity capability at the center of Google’s rollout decision. The company says Argon can find, validate and patch critical software vulnerabilities, and says trusted defenders and Google’s internal teams will receive it without cyber guardrails. That could give security teams a tool for vulnerability work, while the same capabilities make controlled access and careful evaluation consequential.

For businesses considering Argon for coding, legal, finance or other knowledge work, Google’s benchmark claims offer an early indication of intended uses, not a guarantee of performance. The model is not yet described as generally available. Its token pricing and large output limit may matter to organizations planning extended tasks, but the announcement gives no usage examples or costs for typical workloads.

Google’s Phased Model Rollout

Google framed the announcement as an initial, limited release rather than a public launch. It said it is participating in the U.S. government’s voluntary process for pre-release model access while gradually expanding access. The company plans to gather feedback from early testers and revise guardrails before making Argon available to developers, enterprises and consumers.

The announcement also describes internal use as evidence of the model’s potential across engineering and research. For example, Google says Argon helped its quantum computing researchers improve a subroutine’s spacetime resources, beating a published baseline by 40% in one case. These examples come from Google and are not presented as independently audited results. The company says large code rewrites will receive additional review before being deployed.

““Safely releasing frontier capabilities at this level requires a phased approach.””

— Google, in its Sept. 30 announcement

Access, Safeguards and Validation

Google has not given a date for general availability or specified how many organizations will take part in the initial rollout. The announcement does not explain which cyber guardrails will apply to broader releases, how access decisions will be made, or what monitoring will accompany use by trusted defenders. It also does not provide independent confirmation of the reported benchmark scores or internal project results.

The source text mentions Wiz using Argon through its Scan for Good initiative, but the supplied announcement ends before describing the work or its results. Further details about that use cannot be confirmed from the material available here.

Feedback Before Broader Release

Google says it will expand access gradually, collect feedback from early testers and update safeguards before offering Argon to developers, enterprises and consumers. It has not announced a timetable for those steps. The next useful details will include availability dates, access requirements, the safeguards used for each group and further information about the model’s performance outside Google’s reported evaluations.

Key Questions

Who can use Gemini 4 Argon now?

Google says rollout is beginning with trusted cyber defenders through its Fairwind Program, along with its own internal teams. It has not said how many outside organizations have access.

When will Argon be generally available?

Google has not provided a date. It says broader access will follow early testing, feedback and further work on safeguards.

What price did Google announce?

Google listed introductory prices of $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced 95% below the standard input rate. The announcement does not give a general availability date.

Are the reported benchmark results independently verified?

The announcement presents the scores as Google’s reported results. The supplied material does not describe independent replication, so readers should treat them as company-reported evaluations.

Source: hn