Alphabet Inc.’s Google has begun rolling out Gemini 4 Argon, its flagship artificial intelligence model, to a small group of trusted cybersecurity partners. While the model secured leading benchmark scores—including outperforming OpenAI’s Astra in security evaluations—internal employee skepticism has surfaced regarding its real-world performance, particularly in handling complex software coding tasks.
Targeted Release Prioritizes Cybersecurity Partners
Google unveiled Gemini 4 Argon on Wednesday. The initial release targets a restricted audience of trusted cybersecurity partners before the company plans a broader expansion to paid subscribers.
Formidable Metrics Outperform OpenAI’s Astra
On paper, the model boasts formidable metrics. Alphabet claims Gemini 4 posted leading scores across several industry benchmark tests. Notably, it outperformed OpenAI’s Astra on evaluations designed to measure core security skills.
Benchmarks, however, are notoriously slippery.
Internal Skepticism Arises Over Complex Coding Tasks
While the quantitative scores look impressive in a slide deck, internal reception tells a different story. Insiders with direct access to the project report that Gemini 4 struggles when employees put it to work on practical engineering challenges. Specifically, the model stumbles on complex coding tasks that require deep architectural reasoning rather than simple pattern matching.
Internal skepticism remains high as the company prepares its wider rollout.
Staged Deployment Mitigates Risk Before Public Release
Google is proceeding with caution. By restricting initial access to cybersecurity partners, the company is attempting to stress-test the model’s defensive capabilities and catch edge-case vulnerabilities before a public release.
Following this initial validation phase, access will expand to paid subscribers.
