Google Launches Gemini 4 Argon — New Flagship Model, Limited First to Trusted Cyber Defenders

Gemini 4 Argon: Google's New Flagship Model

Google DeepMind has announced Gemini 4 Argon, its next-generation frontier model built for complex, long-horizon tasks: real-world software engineering, enterprise knowledge work (legal research, financial analysis), and autonomous cybersecurity defense. The model features an industry-leading 1 million token output limit — up from 64K previously — letting it reason deeply through complex problems in a single pass.

According to Google, Argon leads on several recognized benchmarks: 77.9% on DeepSWE v1.1 (long-horizon software engineering tasks), #1 on Zapier's AutomationBench (51.3%, measuring end-to-end business task execution), and a leading score on CWE-bench v1 (68%), which evaluates vulnerability remediation.

What stands out most is how it's being released. In contrast to the breakneck model-release race that has defined recent weeks, Google is taking a deliberately gradual, controlled approach. Argon is initially available only to a small group of "trusted cyber defenders" through Google's dedicated Fairwind Program, while the company engages with the U.S. government's voluntary pre-release model access process. Google says it's continuing to strengthen four key safeguard areas — misuse prevention, prompt-injection defense, misalignment monitoring, and hardened execution environments — before opening Argon to developers, enterprises, and consumers.

Introductory pricing is set at $2 per million input tokens and $10 per million output tokens (rising later to $4/$20).

OpenAI Discloses and Disrupts an Attempt to Extract Its Model's "Reasoning"

Separately, OpenAI disclosed that it detected and disrupted a coordinated campaign to extract "protected reasoning" from its models — a technique known as adversarial distillation, where an outside party manipulates a model into revealing its internal reasoning process, which can then be used to train a cheaper competing model. The activity, which began July 1 and spiked to roughly 16,000 requests from 4,000+ users within two days, was partly attributed to individuals associated with Moonshot AI, developer of the Chinese Kimi model (which Anthropic has also previously accused of similar activity). OpenAI emphasized this was not a security breach or access to user data, but manipulation of model interactions, and it shared findings industry-wide through the Frontier Model Forum.

Also in Brief

The U.S. FTC reportedly opened an investigation into OpenAI, Anthropic, and other AI companies over safety risks — according to Reuters, this month's wave of "rogue AI agent" incidents accelerated the decision.

Why This Matters for Small and Medium Businesses

Google's gradual, controlled rollout of Argon — alongside OpenAI's public disclosure of an attempt to extract its proprietary model reasoning — signals an encouraging shift: leading AI vendors appear to be prioritizing transparency and safety testing over a race to ship fastest. For a small or medium business evaluating advanced AI tools, this is a sign that top-tier providers are investing in access control and data protection — a point worth checking with any AI vendor you're considering working with.