Google drops Gemini 3.1 Pro with double the reasoning power

AI benchmark flexes usually feel like background noise at this point, but Google's latest drop is hard to ignore. On ARC-AGI-2, a notoriously brutal benchmark designed to test novel logic patterns, Google's newly released Gemini 3.1 Pro clocked a verified 77.1%. That is more than double the core reasoning performance of its predecessor, Gemini 3 Pro.

Google officially rolled out Gemini 3.1 Pro in preview, promising general availability soon. The release comes hot on the heels of an update to Gemini 3 Deep Think just last week, with Google positioning 3.1 Pro as the baseline engine powering those breakthroughs.

High scores and real-world agent chops

This update is squarely aimed at hard problems where quick autocomplete answers fall flat. Aside from the ARC-AGI-2 jump, Google highlighted major progress on Humanity's Last Exam, and third-party trackers like Artificial Analysis already place the model at or near the top of the pack.

It is also turning heads in practical agentic evaluations. Brendan Foody, CEO of AI hiring and evaluation platform Mercor, noted that Gemini 3.1 Pro immediately grabbed the top spot on their APEX-Agents leaderboard, which tracks performance on real-world professional tasks. In Foody's view, the score underscores just how fast agents are mastering actual knowledge work.

Google designed the model to excel at multi-step problem solving, visual explanations of complex topics, and dense data synthesis across consumer and developer platforms.

Everywhere you work, with no price hike

Perhaps the best news for developers is the price tag: in the Gemini API, 3.1 Pro costs the exact same as Gemini 3 Pro. You get roughly double the reasoning muscle for the same bill.

Google is rolling the preview out across its entire stack at once. Developers can tap it right now via Google AI Studio, the Gemini CLI, Android Studio, and the Google Antigravity agentic platform.

Enterprise teams have access through Vertex AI and Gemini Enterprise. On the consumer side, the Gemini app is offering 3.1 Pro with higher rate limits for Google AI Pro and Ultra subscribers, while NotebookLM is getting the model exclusively for those top two tiers.

Why it matters

We think this release signals that the frontier AI race is no longer about writing better emails; it is about building reliable, autonomous reasoning engines.

For developers and power users, testing 3.1 Pro right now is an easy call. Because API pricing remains unchanged, anyone running multi-turn coding workflows, research synthesis, or autonomous agents gets an immediate performance boost without burning through extra budget. Shipping it straight into NotebookLM is especially clever, turning abstract model smarts into an instant, practical workspace tool.

For the industry, Google is keeping the pressure on OpenAI and Anthropic. Dropping an upgrade this significant just months after Gemini 3 Pro shows that the iteration cycles for production-grade reasoning models are shrinking fast.

The catch? Peak scores on benchmark leaderboards do not always translate smoothly to messier enterprise data pipelines. While topping the APEX-Agents chart looks promising, we think the real test will be whether 3.1 Pro can maintain this level of accuracy when broad traffic hits the API after general availability.