Google DeepMind has reclaimed the top spot on the Artificial Analysis Intelligence Index with the release of Gemini 3.1 Pro Preview. This new iteration demonstrates that high-tier intelligence doesn't always require a premium price tag, as the model matches or exceeds the performance of frontier rivals like Claude Opus 4.6 while remaining significantly more cost-efficient for enterprises.

The model's standout capability lies in its sophisticated reasoning and scientific knowledge. It notably excelled in the CritPt benchmark—a rigorous test involving unpublished, research-level physics problems—surpassing its nearest competitor by five percentage points. For developers, Gemini 3.1 Pro offers top-tier coding performance, leading in tests that measure a model's ability to use a computer terminal like a human programmer (agentic coding).

Perhaps most importantly for reliability, Google has slashed the model's tendency to "hallucinate," or confidently state incorrect information when it doesn't know the answer. By improving its internal knowledge accuracy and implementing better self-awareness of its limitations, the hallucination rate dropped by nearly 40 percentage points. While it still trails slightly in complex real-world multi-step tasks (agentic performance), its blend of speed, multimodal reasoning, and massive 1-million-token context window makes it a formidable tool for high-volume technical applications.

Google DeepMind has reclaimed the top spot on the Artificial Analysis Intelligence Index with the release of Gemini 3.1 Pro Preview. This new iteration demonstrates that high-tier intelligence doesn't always require a premium price tag, as the model matches or exceeds the performance of frontier rivals like Claude Opus 4.6 while remaining significantly more cost-efficient for enterprises.

The model's standout capability lies in its sophisticated reasoning and scientific knowledge. It notably excelled in the CritPt benchmark—a rigorous test involving unpublished, research-level physics problems—surpassing its nearest competitor by five percentage points. For developers, Gemini 3.1 Pro offers top-tier coding performance, leading in tests that measure a model's ability to use a computer terminal like a human programmer (agentic coding).

Perhaps most importantly for reliability, Google has slashed the model's tendency to "hallucinate," or confidently state incorrect information when it doesn't know the answer. By improving its internal knowledge accuracy and implementing better self-awareness of its limitations, the hallucination rate dropped by nearly 40 percentage points. While it still trails slightly in complex real-world multi-step tasks (agentic performance), its blend of speed, multimodal reasoning, and massive 1-million-token context window makes it a formidable tool for high-volume technical applications.