Google limits Gemini 4 Argon access to trusted users only.

Google released Gemini 4 Argon on Wednesday to a curated group of cybersecurity partners. The company calls it its most capable model to date.

• Argon ties with OpenAI's GPT-6 Astra on CWE-bench, a security vulnerability test.
• Google claims Argon sets a new record on real-world, long-horizon software engineering benchmarks.
• 300-plus tebibytes of data center memory freed internally using Argon, without new hardware.
• Gemini 4 arrives nearly a year after Gemini 3, the previous flagship generation.

Why it matters: Google spent the past year releasing smaller, cheaper Flash models while OpenAI and Anthropic pushed further ahead on frontier tasks like complex coding and cybersecurity.

• A promised Gemini 3.5 Pro, due in June, was never released — internally delayed by poor performance, per earlier reporting.

How 18 sources split on this story

Where they split: Coverage agrees on the facts but splits on how much weight to give the benchmark claims versus Google's history of delayed and underperforming frontier releases.

What’s next: Google is working through a U.S. government pre-release safety evaluation before a broader launch.

• Paying subscribers are next in line once the safety process concludes.
• Whether the developer community will judge Argon competitive with OpenAI and Anthropic's current frontier models in real-world use.
• When Google will set a public release date and what the government safety review will require.