Gemini 4 Argon versus GPT-6 and Claude: where does the new Google model excel?

Google recently launched Gemini 4 Argon, an innovative AI model focused on complex tasks such as software development, business analytics, and cybersecurity. During the presentation of this model, Google shared interesting benchmark results demonstrating that Argon can compete with existing models like GPT-6 Astra and Claude Fable 5.1. However, the results reveal no clear winner that is superior in every aspect.

Gemini 4 Argon is initially offered to a select group of cybersecurity partners, and the model is also being utilized by Google itself. A specific date for general availability is not yet known, but the presented benchmark results provide an indication of its performance. Argon achieved the highest score in twelve of the nineteen metrics, showcasing some strong points, particularly in knowledge work and processing longer contexts.

The benchmarks reveal that Gemini 4 Argon has a clear advantage in knowledge work, scoring 68.9 percent on the Vals Index, while Claude Opus 5.5 remains at 67 percent. Additionally, on AutomationBench, Argon scores significantly higher at 51.3 percent compared to its competitors. In financial and legal tests, Argon also performs well, though the percentage in legal tests remains low, highlighting the challenges of the task. However, for software development, the competition is divided, with mixed results where no single model can be deemed the best.

One notable outcome is Argon's strong score in long context processing in the GraphWalks test, where it achieved a score of 84.2 percent. This suggests that Argon performs well in situations where a lot of information must be processed simultaneously. Additionally, the model also scores high in multimodal understanding. While GPT-6 Astra and Claude models have some strengths in various benchmarks, the comparison presents a complex picture. Overall, Argon seems to be a strong contender on paper, but actual performance in everyday use will depend on broader access and practical tests.

Read the full article from AI Insider.