Google's Gemini 4 Argon is closer to the AI frontier than any Google model in months
decoder ·
Google DeepMind's new model, Gemini 4 Argon, has achieved a significant milestone by scoring 97.5 percent on the ARC-AGI-3 benchmark, making it the second lab after OpenAI to cross the 95 percent threshold for general problem-solving. This performance was recorded using the Provider Adapter Harness, which leverages vendor-specific features like retained reasoning and context compaction at a cost of $4,800. In public games, Argon required 62 percent fewer actions than the human baseline and frequently matched speedrun records. However, results vary drastically depending on the testing environment; in the Standard Harness that restricts internal reasoning, the score drops to 49.8 percent. The discrepancy highlights ongoing debates about how API settings influence benchmark outcomes, similar to previous disputes involving OpenAI models. According to Epoch AI, Argon scores 165 points on the Epoch Capabilities Index, placing it just behind leaders GPT-6 Astra and Opus 5.5. While Argon leads in knowledge, agent, and vision tasks, it currently ranks lower in mathematics and software engineering domains. This achievement marks Google's closest approach to the AI frontier since February 2026.
decoder