Gemini 4 Is Here, and Google’s Flagship Tops All Other AI Models on Cybersecurity

What it claims

Google's Gemini 4 Argon tops a 15-try indirect prompt injection test and a first-try exploit-writing test. Its headline coding score, however, comes from a measurement Google ran itself. Argon leads 12 of the 18 benchmarks in Google's own comparison table, ties one and trails on five. The five it trails span coding, science and computer-control tests, but the page does not say which ones.

Evidence

• DeepSWE v1.1 (long, real-world software engineering tasks, % solved). The comparison is not like-for-like: Google computed Argon's 77.9% itself, while rivals' numbers come from a public leaderboard and company reports.
• Claude Opus 5.5: 74.2%
• GPT-6 Astra: 74.1%
• Claude Fable 5.1: 67.4%
• Gemini 3.6 Flash: 49% (July)
• Gray Swan Indirect Prompt Injection. Malicious instructions are hidden in content the agent reads. The score is the attack success rate within 15 tries, so lower is better. The spread matters more for agent builders than Argon's narrow lead.
• Argon: 0.7%
• Claude Opus 5.5 and Claude Fable 5.1: 1.0%
• GPT-6 Astra: 8.5%
• Grok 4.6: 51.8%
• Kimi K3: 52.7%
• Wiz Penetration Test Benchmark. This is an internal Google test. The model must write working exploits against real web-application flaws without seeing the source code, and the score is the share solved on the first try. Argon scored 70.9% against 58.2% for Gemini 3.8 Flash Cyber.
• Anecdotal. Google says Argon helped Wiz find a critical flaw in hospital healthcare software that earlier frontier models had missed. The page gives no details.

How to use it

• Access order: 1. Vetted defenders through the Fairwind Program, Google's limited-access cyber defense initiative. It launched September 2 with more than 650 partners, including governments and critical infrastructure operators. This tier ships "without cyber guardrails," meaning without the refusals that normally block hacking assistance. 2. Paid API customers and Google AI Ultra subscribers, "as soon as possible." No date is given.
• Max output: 1 million tokens per reply, up from 64,000.
• Pricing: introductory rates are $2 per million input tokens and $10 per million output tokens. Standard rates are $4 and $20, which doubles costs. Google has not said when the introductory period ends, so budget against the standard rates.

Limits

• The DeepSWE lead depends on Google's self-computed score, and Argon trails on 5 of 18 benchmarks in Google's own table.
• The Wiz benchmark is internal to Google, so it cannot be checked independently.
• The page does not say whether the general API release will carry cyber guardrails. It says only that the unguarded version goes to Fairwind partners first.
• Google is also taking part in the U.S. government's voluntary pre-release model access process. The page does not describe what that involves.