Gemini 4 Argon opens to cyber defenders before developers

Google released Gemini 4 Argon on September 30, 2026, and calls it its most powerful model yet. Most developers cannot use it yet. Google is giving it first to trusted security teams through a program called Fairwind, and paid API customers come later.

That order says a lot about the model. Google built Argon to find and fix security holes in software on its own. A model that can do that for defenders can also help attackers, so Google is holding back the wide release.

What Google says Argon can do

Argon is aimed at long, multi-step work: coding, research, writing and cybersecurity. Google's announcement says Argon can now write up to 1 million tokens in a single answer, up from 64,000. A token is a small chunk of text, often a short word or part of a longer one. More output room means one request can produce a whole rewritten file, or many of them.

Google says Argon is built to "autonomously find, validate, and patch critical software vulnerabilities," TechCrunch reports. It can also analyze videos and charts. Google staff already use it for debugging and for moving large codebases to new libraries, TechCrunch says.

Google published these scores:

Benchmark	What it tests	Argon's score
DeepSWE v1.1	Real-world software engineering tasks	77.9%
AutomationBench	Automation tasks, where Google says Argon ranks first	51.3%
CWE-bench v1	Fixing known types of security flaws	68%, tied for first
LVBench	Understanding long videos	91.7%

9to5Google compares the DeepSWE result with two rivals. Anthropic's Claude Opus 5.5 scores 74.2% and OpenAI's GPT-6 Astra scores 74.1%. TechCrunch adds that Argon also beats GPT-6 Astra and Anthropic's Fable and Opus models on the Vals AI model index.

Who can use it, and when

For now, only the Fairwind Program has access. Its members are security teams Google trusts to defend systems. Google says paid API customers and Google AI Ultra subscribers come next, followed by wider access.

"Safely releasing frontier capabilities at this level requires a phased approach," said Koray Kavukcuoglu, a senior vice president at Google DeepMind. The company has reason to be careful. In September, Google disclosed that Gemini broke into three companies during a May security test.

Google lists several safety steps for Argon. It tested the model against indirect prompt injection, where an attacker hides instructions inside a web page or file the model reads. It also used red teams, which are testers whose job is to attack the system. 9to5Google reports that Google watches the model's internal activity for signs of misuse and has hardened the sandboxes where it runs code.

What it costs

Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens. After the introductory period, the price doubles to $4 and $20. Cached input tokens cost 95% less than normal input. Caching means the service stores a block of text you send again and again, such as a long document, so you do not pay full price each time.

Argon is not Google's only coding model. In early September, Google shipped Gemini 3.8 Flash, priced as a mid-tier model for everyday coding work.

What Google says it did with Argon inside the company

Google gives three examples from its own systems. Agents running Argon freed more than 300 TiB of memory across Google's data centers, and Google estimates total savings of 500 TiB to 1 PiB. Argon also helped move libgav1, Google's decoder for AV1 video, to new code that runs 2.7 times faster. 9to5Google reports migrations of more than 800,000 lines of code.

What this means for developers

Do not plan a launch around Argon yet. Unless your team is in Fairwind, you cannot call it today. Keep shipping on the model you use now.

Prepare a test set while you wait. Google's scores come from Google. A 3.7-point lead on DeepSWE is real but small, and your own code may not match the benchmark. Collect 20 to 50 real tasks from your backlog, with known good answers, so you can compare Argon fairly on day one.

Think about the 1-million-token output limit. It allows big jobs in one request, such as rewriting a module or writing a full migration. It also allows big bills. One full 1-million-token answer costs $10 at the introductory price and $20 after it. Set output caps in your code before you test.

Design prompts for the cache. With a 95% discount on cached input, put the large, unchanging part first, such as your repository context or API docs. Put the changing question last. That pattern can cut input costs sharply on long agent runs.

If you work in security, watch Fairwind. Google built Argon for finding and patching vulnerabilities, and defenders get it first. Teams that run code scanning or patch programs should follow how Google opens access.