Story
Oct 5, 2026

Google’s Gemini 4 comeback arrives behind a safety gate

Google is presenting Gemini 4 Argon as a long-delayed return to the AI frontier, while its cautious release strategy—and questions raised by reported internal testing—underscore how little a benchmark win alone can settle.

Google’s new flagship arrives after a long and awkward gap. Nearly a year after Gemini 3, the company had continued to ship cheaper Flash models while OpenAI and Anthropic pushed ahead; a promised Gemini 3.5 Pro release, expected in June, never materialized amid reports of delays and poor morale.

On Wednesday, Google unveiled Gemini 4 Argon, claiming frontier-level performance in software engineering, cyber defense, finance and legal work. Chief executive Sundar Pichai called it a model for “complex workflows, cyber defense and software engineering,” saying Google teams were already using it extensively. The company says Argon ties OpenAI’s GPT-6 Astra on a vulnerability-finding benchmark and has helped free more than 300 tebibytes of data-center memory without additional hardware.

Google’s argument is that Argon is not merely a leaderboard entry but a tool for difficult, long-running work. Tulsee Doshi, Gemini’s product lead, described it as “an incredibly well-rounded” model and said its cyber capabilities were “meaningfully important for defenders.” Demis Hassabis framed the controlled debut as a feature, not a retreat: the company would start with governments and trusted cyber defenders through its Fairwind Program before expanding access.

That gate is central to the release. Google is participating in the U.S. government’s voluntary pre-release process and says it has built systems to monitor Argon’s chain of thought, halt risky behavior and reduce the possibility that feedback teaches the model to evade oversight.

But the rollout also leaves the central test unresolved. Bloomberg, cited by Axios, reported that some Google employees found Argon lacking in internal trials—an account Google disputed. Axios noted the broader problem: models can be trained to shine on benchmarks, while practical performance is decided “in the real world.” Argon may be Google’s biggest catch-up bid yet; its restricted release means independent users have not yet been allowed to decide whether it is the comeback Google promises.