AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

Google announced Gemini 4 Argon on Sept. 30, 2026, and said it is rolling the model out to trusted cyber defenders through its Fairwind Program. Google reports strong results in coding and business benchmarks, but the figures are company-reported, and broad availability dates have not been given.

Google announced Gemini 4 Argon on Sept. 30, saying it is beginning a limited rollout to trusted cyber defenders through its Fairwind Program. The company describes Argon as a model for long, complex work in software engineering, business fields such as legal and finance, and cybersecurity; broader access is planned after early testing and further work on safeguards.

Google says Argon is already being used internally by thousands of its employees for specialized coding, research and writing tasks. The company cites several internal examples, including agents that identified memory optimizations across data-center systems. Google estimates those changes could free more than 300 tebibytes of memory once rolled out, with total savings estimated at 500 tebibytes to one pebibyte. Those are company estimates, and the post does not describe an independent audit of the results.

In another example, Google says Argon agents rewrote 32,000 lines of SIMD code in the Rust version of its open-source libgav1 video decoder. The resulting decoder was 2.7 times faster than the existing Rust port and produced identical video output, according to Google. The company also says it is using Argon in larger C and C++ to Rust migrations, including work on the Fuchsia Zircon kernel. It says these changes face automated and manual audits, emulation tests and review before production use.

Google reports a maximum output limit of one million tokens, up from 64,000 for the previous model, to support longer tasks. The company also reports scores of 77.9% on DeepSWE v1.1, 51.3% on Zapier’s AutomationBench and 91.7% on LVBench. Google says Argon leads on the Vals Index and certain finance and legal evaluations. These are benchmark claims reported in the announcement; the post does not provide enough detail here to compare test conditions across systems.

At a glance
announcementWhen: Announced Sept. 30, 2026; limited rollo…
The developmentGoogle announced Gemini 4 Argon, a frontier model entering a limited release for trusted cyber defenders before a wider launch.

A Limited Test for High-Stakes Work

The announcement combines two developments: a model designed for extended, multi-step tasks and a cautious release plan for capabilities Google says can find, validate and patch software vulnerabilities. That pairing matters because the same ability to reason through complex code may help defenders identify weaknesses, while making careful access controls and testing relevant to deployment.

For organizations evaluating AI for engineering or professional workflows, the reported internal examples and benchmark scores offer an early indication of where Google expects Argon to be useful. They are not, by themselves, evidence that the model will produce the same outcomes in other organizations. Performance can depend on the task, data, tools and review process, and Google has not yet described general customer access in detail.

Google says it will offer introductory pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced 95% below the input rate. Those figures give prospective users an initial cost reference, but the announcement does not specify a release date or all the terms that would shape real-world usage costs.

How Google Plans the Rollout

Google says Argon is being made available first to a set of trusted cyber defenders through its Fairwind Program. The company describes that access as part of a phased release, during which it will gather feedback and iterate on guardrails. It also says it is participating in the U.S. government’s voluntary process for pre-release model access. The announcement does not describe the process’s requirements or identify participating government agencies.

Google says trusted defenders and its own internal teams will receive Argon without cyber guardrails so they can use its cybersecurity defense capabilities. Separately, it says the model will become available to developers, enterprises and consumers “as soon as possible,” but gives no firm schedule. The company’s post also names Wiz as a user of Argon through its Scan for Good initiative, described as a program for cybersecurity defense.

The company frames the model’s long output limit as support for sustained reasoning and complex workflows. Google’s examples range from code optimization to research across documents and analysis of long videos. That framing explains the breadth of the announcement, but each example remains a Google account of the model’s performance rather than an independently verified finding.

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

— Google

Release Dates and Results Still Open

No date for general availability is included in the announcement, and Google does not say how many testers are in the initial group or which organizations will receive access. The company has not set out a schedule for expanding the rollout or detailed what changes to guardrails will be required before that happens.

The performance figures and operational examples come from Google. The announcement does not provide independent evaluations of the claims, full benchmark methodology, or enough information to assess whether reported gains will carry over to other users’ environments. Google also does not quantify how often Argon identifies vulnerabilities correctly, how its patches are reviewed, or what limits apply to the cyber capabilities made available to testers.

Feedback Before Wider Access

Google says it will use feedback from early testers to refine Argon’s safeguards while gradually expanding access. The next clear milestone is a broader release to developers, enterprises and consumers, but the company has not announced when that will happen or specified which products will include the model.

Further information from Google about rollout timing, access terms, safeguards and evaluation methods would help prospective users judge the model’s availability and reported performance. Until then, Argon remains in a limited testing phase, with the company’s announcement serving as the source for its capabilities, benchmarks and internal results.

Key Questions

What is Gemini 4 Argon?

Google describes Gemini 4 Argon as a frontier model for complex, long-running tasks in software engineering, enterprise knowledge work and cybersecurity defense.

Who can use Argon now?

Google says the model is rolling out to trusted cyber defenders through its Fairwind Program and is also being used by internal teams. The announcement does not name all participants.

When will Argon be available to everyone?

Google says it plans to make Argon available to developers, enterprises and consumers as soon as possible, but has not given a date.

What does Google say Argon costs?

Google lists introductory prices of $2 per million input tokens and $10 per million output tokens. Cached input tokens are priced 95% below the input rate.

Are Argon’s reported results independently verified?

The cited benchmark scores and internal project results are reported by Google. The announcement does not include independent verification or full evaluation details.

Source: hn

FALL

Fall Picks

As an affiliate, we earn on qualifying purchases.

You May Also Like

AI-Powered Ideas For Halloween Decorating

AI Halloween decorations can generate content or respond to visitors, but some use only sensors and recordings. Check features, privacy and reliability.

Jun Kim, oMLX Creator And Maintainer, Joins Hugging Face To Support The MLX Community

Hugging Face says oMLX will receive funded maintenance under creator Jun Kim, while keeping its Apache 2.0 license and existing leadership.

Maximize Efficiency With AI Tools & Automation In 2026

Explore how AI tools and automation are transforming productivity and security in 2026, with confirmed developments and key insights for users.

DraftKings Is Using AI To Behaviorally Target Chronic Gamblers

The EFF says DraftKings uses betting records to identify customers likely to lose and sends them targeted promotions to encourage more betting.