Google unveils Gemini 4 Argon with 1 million-token output

The video highlights three technical evolutions. First, Argon can generate up to 1 million output tokens in a single request, far beyond the tens of thousands token limit of most frontier models. Second, the model has been tuned to eliminate “AI slop,” delivering concise, precise prose with reduced hallucinations. Third, Argon’s autonomous agents persist beyond a single instruction, researching, designing, and implementing tasks such as building a toddler reading app without stopping.

Google says the development cycle has been accelerated by internal agents and large‑scale dog‑fooding by roughly 100,000 employees. To guard against over‑reliance on AI, the team proposes “verifying AI with AI” and sharing prompts for scrutiny. Integration with simulators and robotics is already underway, pointing to broader physical‑world deployments.