Gemini 4 Argon: 1M output, cyber-first rollout
Google's Argon starts with trusted cyber defenders. We examine its 1M output limit, access, pricing and vendor benchmarks.
View the original post ↓Product claims checked against the official documentation linked below. Vendor benchmarks are attributed to the vendor.
On 30 September 2026 Google posted that it was introducing Gemini 4 Argon. The English line, which the screenshot translates, is:
Today we’re introducing Gemini 4 Argon. It delivers frontier performance in complex workflows across real-world software engineering, knowledge work, and cybersecurity defense with an industry-leading 1M token output limit.
The source of record is the Keyword post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google. The Spanish in the feed is X’s translation.
This note separates the tweet, who can actually call the model, the price, and the benches Google published. It is not a bake-off from this desk.
What Argon is, in Google’s words
Argon is the first model released under the Gemini 4 name. Google describes it as built to sustain deep reasoning across long-horizon workflows: real-world software engineering, enterprise knowledge work (legal, finance), and cybersecurity defense.
The technical change that will travel is the output window. Previous Gemini models topped out at 64K output tokens. Argon goes to 1 million. Google’s line is that headroom to think and emit hundreds of thousands of tokens in one trajectory changes how hard a problem you can finish in a single run.
That is an output limit, not a claim that every answer will be a million tokens. A 1M ceiling also means a 1M bill if a loop actually uses it. At the introductory output price below, a full million-token completion is $10 before the intro ends, and $20 after.
Who gets it on day one
Not the Gemini app. Not a public API key.
Google is rolling Argon out first to trusted cyber defenders through the Fairwind Program. The company says it is in the U.S. government’s voluntary pre-release process, and that it will gather feedback and iterate on guardrails before a wider release.
The planned next door, with no date on the announcement: paid API customers and Google AI Ultra subscribers, then developers, enterprises and consumers “as soon as possible.”
For that trusted cohort, Google says it will release Argon without cyber guardrails, so defenders can use the full defensive capability. That sentence is the product decision. It is also why the public door is closed. A model trained to find, validate and patch vulnerabilities is not something you drop into a free chat box on day one.
Google says Argon can autonomously find, validate and patch critical vulnerabilities. On an internal vulnerability benchmark it reported exposures across codebases in 20 languages. On a Wiz black-box test (live web systems, no source) it outperformed 3.8 Flash Cyber. In an early demo, Google says it found a critical vulnerability exposing personal data in healthcare software used by hospitals, a case previous frontier models missed. No public CVE is in the announcement. Treat that as a vendor case study until a disclosure exists.
Price: the sticker and the footnote
| Introductory | After the intro | |
|---|---|---|
| Input | $2 / million tokens | $4 / million |
| Output | $10 / million tokens | $20 / million |
| Cached input | 95% off the input price | not restated in the footnote |
The footnote does not name the end date of the introductory window. Confirm it on the live pricing page before you budget a quarter around $2/$10.
The post-intro row ($4 / $20) matches the list price Anthropic published for Opus 5.5. The intro row matches Sonnet 5.5’s list price. That comparison is arithmetic, not a quality claim.
The benches Google put on the page
All of these are Google’s figures, or leaderboards Google says it leads. Not measured here.
| Eval | Argon, as published | What it is |
|---|---|---|
| DeepSWE v1.1 | 77.9%, called state of the art | Long-horizon software engineering |
| AutomationBench (Zapier) | 51.3%, rank #1 | End-to-end business tasks |
| LVBench | 91.7%, called state of the art | Long video understanding |
| CWE-bench v1 | 68%, tie for first | Vulnerability remediation |
| Vals Index | leading; secondary coverage cites 68.9% | Finance, coding, legal, tax, GDP-weighted |
Secondary write-ups the same day put Opus 5.5 at 74.2% and GPT-6 Astra at 74.1% on DeepSWE v1.1, and Opus 5.5 at 67.0% / Astra at 63.1% on the Vals Index. Those competitor numbers are not in the sentences I pulled from the Google post itself. If you cite them, cite the recap, not Google’s mouth.
Google also says Argon leads 12 of 18 benchmarks in its comparison table against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1, and ties one. It trails on several coding, ML-engineering, science and computer-use tests. A table where you win most cells and lose the ones closest to your daily job is not “best model.” It is a workload split.
What I would not do with the announcement
- Do not tell a client Argon is in their Gemini account. On 30 September it was not.
- Do not plan a cyber workflow around “no guardrails” unless you are actually in Fairwind. That mode is for the trusted cohort.
- Do not budget the intro price as permanent.
- Do not read 1M output as “the context window is new.” The headline change is output, up from 64K.
Frequently asked questions
Is Gemini 4 Argon generally available? Not on announcement day. Fairwind first. Wider access starts, in Google’s plan, with paid API and AI Ultra. No date.
Did the 1M figure replace the context window? Google’s sentence is an output-token limit, up from 64K. Do not conflate it with input context unless the model card says so.
Is the $2 / $10 price permanent? No. Footnote: after the introductory period, $4 in / $20 out. End date not in the post.
Is this a public cyber tool? The defensive capability is the reason for the phased door. Unguarded cyber use is described for trusted defenders and internal Google teams, not for a consumer chat.
Takeaway
Argon is a frontier Gemini with a million-token output ceiling and a cyber-defense brief, priced like a mid-tier model only for an unnamed intro window. The tweet is real. The public product is not open yet.