Gemini 4 Argon Launches, but Only Through Fairwind
Google announced Gemini 4 Argon on September 30, and most of you can't touch it yet. Access starts with a set of trusted cyber defenders in Google's Fairwind Program, so the practical question for a leader is what to do with a frontier model you can read about but not call.
What happened with Gemini 4 Argon
Google describes Argon as a frontier model for coding, enterprise knowledge work and cyber defense. In its launch post, Google says it is "rolling out to a set of trusted cyber defenders through our Fairwind Program." TechCrunch reports the model was trained for defensive work and can "autonomously find, validate, and patch critical software vulnerabilities."
What Google has published so far:
- Access: Fairwind participants first. Broader access is promised to "developers, enterprises, and consumers, starting with paid API customers and Google AI Ultra subscribers."
- Guardrails: Fairwind users get Argon without cyber guardrails, for defensive work. Wider audiences get a version with safeguards against misuse, prompt injection and misalignment.
- Pricing: $2 per million input tokens and $10 per million output tokens at launch, with cached inputs at a 95% discount. After the introductory period: $4 and $20.
- Output limit: 1 million tokens, which Google calls industry-leading.
- Self-reported benchmarks: 77.9% on DeepSWE v1.1, 68% on CWE-bench v1 (vulnerability remediation), 51.3% on AutomationBench.
TechCrunch adds that Google claims Argon beats OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models on various benchmarks. Those are vendor numbers, so treat them as a reason to run your own tests, not as a verdict.
What is missing: a general availability date. Google says only "rolling out soon." Treat the pricing as the introductory deal, not the long-term number.
Why the Fairwind-only rollout matters for leaders
The gating is the story. A model built to find and patch vulnerabilities is also one that could find them for the wrong people, so Google is starting with defenders. Expect more frontier launches to arrive in tiers instead of all at once.
That changes how you plan:
- Don't put Argon on a committed roadmap date. No GA date means no dependency.
- The public version will differ from the Fairwind one. The guardrails that come with general access may change behavior on security tasks, so don't assume the demos you read about transfer to your account.
- Budget with the higher price. Plan around $4 and $20 per million tokens, and count a 95% cached-input discount only once you have confirmed it applies to your workload.
- Security teams are the first audience. If your company works in a Fairwind-type segment, such as critical infrastructure, ask your Google contact about access now.
What to do on Monday morning
You can prepare without the model:
- Write a bake-off plan. Pick three or four real tasks: a codebase migration, a vulnerability triage queue, a long document drafting job. Define pass criteria now, so you are not swayed by launch benchmarks later.
- Keep your model layer swappable. If your apps call one vendor directly, put a thin routing layer in front so trying Argon on general availability day is a config change, not a project.
- Ask your security team where it hurts. Backlog of unpatched findings is the obvious test case for an automated patching model. Agree on human review rules before any model proposes a fix.
- Test long outputs. A 1 million token output limit is only useful if your pipeline can handle it. Check timeouts, streaming and cost caps today.
- Set a review date. Revisit when Google announces paid API access, and decide in advance who owns the evaluation.
Wait for the data, not the headline
Benchmarks come from the vendor, access is narrow, and the version most teams will get is not the one defenders are using. Waiting a few weeks for independent results costs little.
For anyone deploying AI in production, the useful move is to be ready to test Argon the day it opens, and not to promise it to anyone until it does.