What is Gemini 4 Argon?

It is Google DeepMind's new flagship model, announced on September 30, 2026. Google describes it as a model for real-world coding, enterprise knowledge work (for example legal and finance) and cyber defense.

When will Gemini 4 Argon be available?

For now it is reachable only by a group of trusted cyber defenders through the Fairwind Program. Google speaks of broad availability as soon as possible, starting with paying API customers and Google AI Ultra subscribers, without giving a date.

How much does Gemini 4 Argon cost?

At the introductory price, 2 dollars per million input tokens and 10 dollars per million output tokens (cached input at 0.10 dollars). After the introductory period prices rise to 4 and 20 dollars per million tokens.

Does Gemini 4 Argon beat GPT-6 Astra and Claude Opus 5.5?

According to figures released by Google, it leads 12 of 18 benchmarks and ties one, for example DeepSWE v1.1 (77.9% against 74.1% for Astra and 74.2% for Opus 5.5). It trails on FrontierSWE v2 and Terminal-Bench Science. These are company-reported numbers, not yet independently verified.

Why is the release limited?

Google says it is taking part in the voluntary pre-release access process promoted by the US government and is strengthening safeguards against cyber and CBRN (chemical, biological, radiological and nuclear) misuse.

Google Gemini 4 Argon: Benchmarks, Pricing and Availability

On September 30, 2026, Google introduced Gemini 4 Argon, the first flagship in the Gemini family since Gemini 3, released in November 2025. For months OpenAI and Anthropic kept shipping models at a fast pace while Google stood still: Argon is the answer, with numbers that on paper take back the lead on benchmarks. There is one big limit, though: for now almost nobody can use it.

What it is built for

Google describes it as a model for three areas:

Coding on real projects, including code migrations and algorithm optimization
Professional work in business settings, such as financial research, drafting legal documents and analyzing charts
Cyber defense, meaning finding and fixing critical vulnerabilities autonomously

Among the stated features is also a limit of 1 million tokens for long, multi-step reasoning.

The benchmarks: where it wins and where it loses

Among the 18 benchmarks Google disclosed, Argon leads 12 and ties one. A few comparisons with GPT-6 Astra and Claude Opus 5.5:

Test	Gemini 4 Argon	GPT-6 Astra	Claude Opus 5.5
DeepSWE v1.1 (real-world coding)	77.9%	74.1%	74.2%
AutomationBench (business tasks)	51.3%	41.4%	42.5%
Vals Finance Agent v2	65.4%	53.5%	58.6%
Harvey Legal Agent	19.6%	5.4%	3.8%
CWE-bench v1 (fixing vulnerabilities)	68%	68%	67%

Argon does not win everywhere: on FrontierSWE v2 Astra is ahead (65.5% against 55%), and the same goes for Terminal-Bench Science 0.1 (68.1% against 57.6%). All of this is data released by Google, so it should be read as such: the real test will be whether the advantages turn into reliable behavior in everyday use.

Pricing

At the introductory price Argon costs 2 dollars per million input tokens and 10 per million output tokens, with cached input at 0.10 dollars. After the introductory period prices will rise to 4 and 20 dollars. For comparison, GPT-6 Astra costs 10 and 50 dollars: while the introductory offer lasts, Argon costs about a fifth as much.

Why access is limited

Argon is currently available through the Fairwind Program, reserved for trusted cyber defenders. Google explains that it is taking part in the US government's voluntary pre-release access process and strengthening safeguards against cyber and CBRN misuse (chemical, biological, radiological and nuclear).

Broad availability is planned "as soon as possible", starting with paying API customers and Google AI Ultra subscribers, but Google has not given a date.

How Google already uses it

The company says Argon is already working on internal workflows, with results that include:

migrating over 800,000 lines of the Fuchsia Zircon kernel from C/C++ to Rust
libgav1 video decoding that is 2.7 times faster
more than 300 TiB of memory freed in data centers
a quantum optimization 40% better than the published baseline

What to expect

The competitive picture changes, but not entirely: Argon is ahead in many tests, not all, and for API users the real comparison can only start once the model opens up. In the meantime, keep an eye on pricing, because with low introductory prices the competition between Google, OpenAI and Anthropic is pushing down the cost of top-tier models.