The September 30 announcement

Google announced Gemini 4 Argon on September 30, 2026, in a post signed by Koray Kavukcuoglu, senior vice president at Google DeepMind and chief AI architect at Google. Argon is the company's next frontier model, built for long, multi-step work in software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense.

It is not on general sale. Google says Argon is "rolling out to a set of trusted cyber defenders through our Fairwind Program," and describes the release as phased: the company is taking part in the United States government's voluntary process for pre-release model access and says it will widen access gradually before reaching developers, enterprises and consumers. Most flagship models arrive with a model name, a rate card and a context limit. This one arrives with the name of an access program.

Who can use Argon today

Access runs through the Fairwind Program, which Google DeepMind describes as early access for high-priority defenders. Its page says Google currently works with more than 650 partners globally, and that a set of those partners gets exclusive access to Gemini 4 Argon, either as a standalone model or inside CodeMender, Google's code security agent.

Eligibility is narrower than the phrase "cyber defenders" suggests. Google lists governments and national cyber authorities, operators of critical infrastructure in healthcare, telecommunications, energy and financial networks, and core technology platforms. Academic labs that work on defensive benchmarking can apply. Organizations that do not qualify are pointed to CodeMender running publicly available models instead. Google says Wiz is already using Argon through its Scan for Good initiative, which hunts for high-risk exposures in public infrastructure.

Partners accept conditions in return. They must use user-level authentication and phishing-resistant multi-factor authentication, may grant Argon access only to internal cybersecurity, incident response or penetration testing teams, and must track who uses it. Google says partners may not share, redistribute or sell that access, that it runs background checks on applicants, and that the program is limited to dual-use tasks such as authorized threat simulation, reverse engineering and malware analysis. Google also says it will release Argon to those defenders, and to its own internal teams, without the cyber guardrails.

A field of blank geometric tiles approaches a narrow lit gate, with a few blank tiles visible beyond it.
Concept illustration of Fairwind's limited first access. The shapes do not represent specific partners or their number, and the image does not depict public developer access.

How Argon compares with existing Gemini models

Google says it expanded Argon's output token limit to 1 million tokens, up from 64,000 in the previous generation, a figure the company describes as industry-leading. That limit covers how much the model can write in one trajectory, and the announcement does not publish an input context window.

The price carries a footnote that the main text leaves out. Argon "will launch at an introductory price" of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price. Google does not say when the introductory period ends. The footnote on that sentence gives the rest of the schedule: $4 per million input tokens and $20 per million output tokens after the introductory period.

For scale, the Gemini Developer API pricing page lists the newest model in its own list, Gemini 3.8 Flash, at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, rising to $1.50 and $7.50 from January 1, 2027. Against the 2026 Flash rates, Argon's introductory price is about 2.7 times higher and its post-introductory price about 5.3 times higher. Against the 2027 rates the gap narrows, so the comparison depends on which date you use.

What Google's own comparison table shows

Google published a comparison table that puts Argon against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5.

Argon leads on knowledge work: 68.9% on the Vals Index against 63.1%, 65.8% and 67.0%; 51.3% on AutomationBench; 19.6% on Harvey's Legal Agent Benchmark against 5.4%, 6.7% and 3.8%. It also leads on long-context retrieval, multimodal understanding and CWE-bench v1, where it ties GPT-6 Astra at 68%.

Several rows run the other way. Argon trails GPT-6 Astra on FrontierSWE v2 at 55.0% against 65.5%, on Terminal-Bench Science 0.1 at 57.6% against 68.1%, and on OSWorld-2.0 at 69.2% against 72.6%. It trails Claude Opus 5.5 on Terminal-bench 4.0 at 57.4% against 66.4% and on PostTrainBench at 45.3% against 49.3%.

The methodology published with the table explains how to read those figures. Argon's scores are pass@1 with single attempts, no majority voting and no parallel test-time compute, run through the Gemini API at the highest thinking settings unless noted otherwise. The rival figures are not measured by Google: they come from the other providers' self-reported numbers, public leaderboards or system cards. Every number in this section is a vendor-reported score rather than an independent measurement, and Google's own results include a real-world vulnerability set run in a harness the methodology describes as not cyber specialized.

What the developer documentation shows

As of October 12, 2026, none of it mentions Argon. The Gemini API release notes end at October 8, when Google deprecated two Flash model strings, and October 6, when Nano Banana 2.1 reached general availability. The API's model list runs through the Gemini 3.8 family with no Argon entry, and the pricing page has no Argon row.

Without a published model ID, no developer can call Argon through the Gemini API today, and the $2 and $10 figures cannot be checked against a rate card.

What to watch next

Google says the next group to get access will start with paid API customers and Google AI Ultra subscribers, and it has not dated that step. A model ID and a pricing row on the developer documentation would mark the change from a defense program to a product developers can buy. Two questions stay open: how long the introductory price lasts, and what the input context window is.

A security team at a hospital, a telecom operator or a national cyber authority can apply through Fairwind, where the gate is vetting rather than budget. Everyone else is waiting for the model list and the rate card to catch up.