Frontier AI · AI models

Gemini 4 Argon: Google’s Big AI Upgrade Starts With Cyber Defenders

A larger output budget could help AI tackle longer assignments. The immediate questions are who gets access, what it costs and how teams will check the work.

The Super Intelligence Times · September 30, 2026 · Reporting & analysis · 4 min read

Editorial illustration · AI-generated. A conceptual scene, not a photograph of Google hardware.

Google announced Gemini 4 Argon on September 30, with an output ceiling of one million tokens and an initial rollout to selected cyber defenders. Wider availability will follow. Google’s announcement describes a phased release.

For businesses evaluating AI agents, this raises a practical question: can a model carry a difficult assignment through to a result that survives review? A longer response is useful only when the extra work improves the outcome. That is the test we would put ahead of a launch-day ranking.

What a million-token output limit changes

Google says the ceiling rises from 64,000 tokens to one million, allowing longer reasoning and generation within a single run. This is an output limit, not an announcement of a one-million-token input window.

Our reading: the useful scenario is a job with many dependent steps. Consider a software migration that needs an initial plan, code changes, tests and revisions after failures. Extra room may help an agent keep working through that sequence. It still needs a clear definition of completion and a way to check each result.

A buyer should therefore ask how much work reached an acceptable standard, how many corrections a reviewer made and what the complete attempt cost. Measuring generated text alone would reward the wrong thing.

Read the pricing footnote

Google’s quoted prices per one million tokens
Token type	Introductory	After introduction
Input	$2	$4
Output	$10	$20

Google also lists a 95% cached-input discount, equivalent to $0.10 per million at the introductory input rate. Its pricing footnote supplies the later rates but no introductory end date.

For illustration, 100,000 billed input tokens plus 20,000 billed output tokens would cost $0.40 at the introductory rates, or $0.80 at the later rates. That calculation assumes uncached input and excludes other charges. It is arithmetic from the listed rates, not a measured cost for completing a business task.

Canadian teams should confirm billing currency, taxes and account terms before budgeting. For a pilot, set a spending cap for the whole assignment, including retries. Track review time alongside the token bill: an inexpensive answer that requires extensive repair can still be expensive work.

Who can use it first?

The Fairwind Program gives selected partners access to Argon for security work, including through Google’s CodeMender agent. DeepMind describes priority defenders such as governments, healthcare providers and telecommunications services. An application is subject to review; submitting one is not an access guarantee.

The program’s rules are specific. Participating organizations must use controls including phishing-resistant multi-factor authentication, restrict Argon access to internal security teams and track employee use. They cannot share, resell or redistribute access. Permitted dual-use activity is limited to defensive and academic research purposes.

Google plans broader distribution starting with paid API customers and Google AI Ultra subscribers. The launch post gives no fixed release date. It should not be read as confirmation that ordinary subscribers can select Argon today.

Our take: prepare a test that can fail

An Alberta software team can prepare a useful evaluation before it gets access. Choose a past, non-sensitive task with a known outcome. Give the model the same constraints a colleague had, keep its permissions narrow and record what happens when the first approach fails.

For a coding task, require passing tests and a reviewable change. For research, require traceable sources and mark unsupported claims as failures. Compare the finished result with the tool already in use. A polished explanation should not rescue an incorrect deliverable.

The next details to watch are the public access date, complete billing terms and repeatable evaluations outside the launch material. Those will help buyers decide where Argon earns a place in their workflow.

Sources and editorial note

The announcement, expanded pricing footnote and Fairwind Program page were checked on September 30, 2026. Product statements are attributed to Google. We have not independently tested Argon. Practical evaluation suggestions are our editorial analysis. The cover was generated with AI.