Argon Falls Short: Google's New AI Model Raises Questions

Google's latest AI model, Gemini 4 Argon, has been hailed as its most powerful yet. However, according to some employees who have used it, Argon may not be living up to expectations in real-world use.

Bloomberg reports that while Argon excels on benchmarks in areas such as long-context coding and enterprise-agent tasks, it struggles with some coding tasks, particularly front-end design. Employees also claim that the model is so large it may be costly to run, and shows signs of 'benchmaxxing,' or tuning for test scores over practical usefulness.

Google has responded to these claims, stating it is 'inaccurate' to say Argon underperforms at coding. The company's Fairwind Program allows selected cybersecurity partners and Google staff access to the model, which is currently restricted due to its power consumption requirements.