Google Drops Cyber‑Ready Gemini 4 Argon While Rivals Keep Getting Played 🤖

Google released Gemini 4 Argon on Wednesday, positioning the model as its frontier offering for coding, office work and cyber defense. The launch arrives one week after Anthropic's Claude Opus 5.5 and one day after GPT 6.1 Sol, intensifying competition among American AI labs. On DeepSWE v1.1, which scores an AI's ability to finish real software engineering tasks, Argon reached 77.9%. Claude Opus 5.5 scored 74.2%, GPT-6 Astra 74.1% and Claude Fable 5.1 67.4%; Gemini 3.6 Flash managed 49% on the same test in July. Argon also expanded its single-response output to as many as 1 million tokens, up from 64,000, a jump Google's own materials describe as roughly 750,000 words versus about 48,000. Google said Argon leads on 12 of 18 benchmarks, ties one and trails on five, covering coding, science and computer-control tests, while noting its DeepSWE score was self-computed against rival figures drawn from a public leaderboard and company reports.

The model's signature feature is its cyber capability. On Gray Swan's Indirect Prompt Injection benchmark, which hides malicious instructions in content an agent reads and scores how often attacks succeed within 15 tries, Argon posted 0.7%, lower than Claude Opus 5.5 and Claude Fable 5.1 at 1.0% each, GPT-6 Astra at 8.5%, and Grok 4.6 and Kimi K3 at 51.8% and 52.7% respectively. Google is distributing Argon to vetted security teams through the Fairwind Program, a limited-access cyber defense initiative launched September 2 with more than 650 partners including governments and critical infrastructure operators. The company said the model ships "without cyber guardrails," the built-in refusals that normally block hacking assistance, on the reasoning that defenders need attacker-like capabilities to patch vulnerabilities before criminals exploit them; Google is also participating in the U.S. government's voluntary pre-release model access process.

Argon's cyber performance improves on Gemini 3.8 Flash Cyber, the restricted model Google introduced at Fairwind's launch. The approach mirrors rival programs: an early version of Anthropic's Claude Mythos helped identify 271 vulnerabilities in Firefox that Mozilla subsequently patched, and OpenAI runs a comparable Trusted Access for Cyber program. Google's phased rollout is intended to manage the dual-use risk that the same capability could be turned against defenders, with access expanding as safeguards and partner vetting scale.