Google's 'Gemini 4 Argon' Launch Exposes Big Tech's Dual Strategy: Performance Rivalry Meets Security Control

Generative AI companies are increasingly adopting a "staged release" approach, providing cutting-edge models to security experts before making them publicly available. As AI advances beyond merely identifying software vulnerabilities to actually designing and executing attacks, this strategy is seen as an effort to buy defenders time to respond.

On the 13th of this month (local time), Google expanded access to its top-tier AI model "Gemini 4 Argon" to general paid users, developers, and enterprise customers. The model is designed to solve complex, multi-step tasks across coding, finance, law, cybersecurity, and other domains.

Prior to the public launch, Google provided Argon on the 30th of last month to select experts participating in its cybersecurity collaboration program "Fairwind" and to the U.S. government. The move reflected concerns that the model's vulnerability-detection capabilities could be repurposed as hacking tools. During this period, the company said it strengthened the public release version's resistance to "prompt injection" attacks, where malicious commands are used to manipulate AI.

This staged release is not unique to Google. Anthropic in April of this year released "Claude Mythos Preview," a preview version of its cutting-edge "Mythos" model, to only a select group of companies. At the time, the company explained that it judged the technology could be exploited by hackers and therefore prioritized providing it to major technology and security firms for defensive purposes. In June, Anthropic released "Fable5," a safely refined general-purpose version of the Mythos-class model, and "Mythos5," a security-specialized model, through its cybersecurity collaboration initiative "Project Glasswing."

OpenAI also adopted a similar approach in September, first providing "GPT-6 Astra" to select organizations participating in its cybersecurity program "Daybreak" before gradually expanding access to paid users over several days. Astra is the first model to receive a "critical" risk rating—the highest level—in OpenAI's own evaluations. It is reportedly capable of identifying security vulnerabilities and designing and executing attack sequences without human intervention, given certain access privileges.

Performance competition is a separate matter... Argon tops evaluation benchmarks

If staged release is a security measure, the general launch that follows is the starting gun for performance competition. According to Google DeepMind, Argon achieved sole first place in 13 of 19 evaluation categories compared against competing models.

On "AutomationBench," which measures work automation capabilities, Argon scored 51.3%, ahead of OpenAI's GPT-6 Astra (41.4%) and Anthropic's "Claude Opus 5.5" (42.5%). On "DeepSWE," which evaluates long-duration software development capabilities, it scored 77.9%, higher than GPT-6 Astra (74.1%) and Claude Opus 5.5 (74.2%).

Independent evaluations showed similar results. On the "Vals Index," a professional task performance assessment by AI model evaluator Vals AI, Argon scored 68.9%, ranking first among 48 models evaluated. Claude Opus 5.5 (67.0%) and GPT-6 Astra (63.1%) followed. On the text aggregate rankings released on the 13th by AI model evaluation platform Arena, Argon also topped the list with 1,525 points.

Google also highlighted internal adoption results. Koray Kavukcuoglu, Senior Vice President and Chief AI Architect at Google DeepMind, said, "We are using Argon internally at Google to solve our most challenging engineering problems." According to Google, after applying Argon to internal AI development tools, code submissions using AI agents increased by more than 35%, while rollback rates—reverting written code to a previous state—decreased by approximately 30%. The Chrome development team used Argon to refactor approximately 6 million lines of the 40-million-line Chromium codebase, completing work that was expected to take months in just a few weeks.

However, external observers note that further validation in enterprise environments is needed. Nitish Tyagi, Senior Analyst at market research firm Gartner, told IT publication Computerworld that "whether [Argon] can deliver commensurate quality of output in enterprise use cases remains unproven." He added that companies should prioritize results from direct testing in their own work environments when selecting AI models.

Meta's entry reshapes the competitive landscape

Meta has also joined the race. On the 12th, Meta unveiled its latest AI models "Muse Wonder" and "Muse Wonder Full." Following last year's performance controversy surrounding Llama 4, Meta recruited Scale AI founder Alexandr Wang and undertook a major reorganization of its AI division. The company introduced "Muse Spark" in April and launched the personal AI agent "Muse" in September, accelerating new product development.

According to independent AI evaluation organization Artificial Analysis, Muse Wonder Full's aggregate intelligence index stands at 56 points, up 8 points from the previous Muse Spark 1.3 (48 points). Its TerminalBench 4.0 score, which evaluates the ability to execute commands and solve problems in real development environments, also rose from 33% to 57%.

User acquisition has been rapid as well. According to mobile app analytics firm Sensor Tower, the personal AI agent "Muse," launched on the 8th of last month, surpassed 5 million cumulative downloads in the U.S. within 22 days—faster than OpenAI's ChatGPT (56 days) and Anthropic's Claude (492 days). IT publication The Verge noted that while OpenAI and Anthropic appeared to be leading the AI model performance race until recently, the situation has shifted as Google and Meta have successively introduced high-performance models.

Staged release is not a sufficient condition for 'safety'

Experts caution that staged release does not guarantee safety. Hwang Seok-jin, professor of cyber forensics at Dongguk University's Graduate School of International Information Protection, explained, "As AI has advanced to the point where it can automate everything from vulnerability discovery to attack execution, companies are trying to secure time to build defensive systems before public release."

He emphasized, however, that "safety is not guaranteed merely by releasing with a time lag; user validation, access control, and real-time exploitation monitoring must all be implemented together." Big Tech's deliberations over how and when to release high-performance AI models are expected to deepen as performance competition intensifies.