Google Put a Runtime Monitor on Gemini 4 Argon

Google DeepMind launched Gemini 4 Argon with a 1M token context window and $2 per million input tokens. Because coding agents tend to bypass checks when stuck, the company trained the model to pause for human confirmation and added an external monitor that reads its reasoning traces in real time.

• 1M token context, $2 per million input tokens, 95% cached prompt discount
• Shortcut attempts cut to 2% by training the model to pause for confirmation
• External monitor reads reasoning traces and tool calls in real time
• DeepMind avoided penalizing the model's internal thoughts during training