Google unveils Gemini 4 Argon — and immediately locks it down

Google has revealed its latest frontier AI model, Gemini 4 Argon. According to chief AI architect and Google DeepMind SVP Koray Kavukcuoglu, the model delivers "frontier performance in complex workflows" — spanning real-world software engineering, enterprise knowledge work in law and finance, and cybersecurity defense.

But access is deliberately restricted at launch. Only a "set of trusted cyber defenders" will get it first. Kavukcuoglu says Google is "actively engaged in the U.S. government's voluntary process for pre-release model access" while it gradually widens availability.

Caution over a splashy launch

The reason for the restraint is a fear the model could turn out to be misaligned. Before any broader rollout, Google promises to strengthen "critical frontier safeguards": defending against misuse, guarding against prompt injection attacks, and monitoring for misalignment.

Inside the company, though, the model is already at work. Kavukcuoglu notes that Gemini 4 Argon is powering Google's "internal workflows," and has helped with tasks such as large-scale codebase migrations.

Google also shared a sizable chart of benchmarks in which Gemini 4 outperforms competing models from OpenAI and Anthropic. Notably, earlier in the week the AI community on X was already buzzing over what appeared to be leaked benchmarks for the new model.

A race that isn't slowing down

The announcement lands just a day after OpenAI's splashy DevDay, where the company launched its Dots AI agent and the GPT-6.1 Sol model. Tellingly, OpenAI said that same week it would not ship a planned GPT-6.1 Astra model — also over safety concerns.

Gemini 4 Argon's arrival also follows a recent shift for Kavukcuoglu himself: he took over as head of DeepMind in August.

Source: The Verge