First impressions of Google Gemini 3.1 Pro: ‘Deep Think Mini’ with customizable thinking on demand

For the past three months, Google’s Gemini 3 Pro has held its ground as one of the most powerful borderline models available. But in the fast-moving world of AI, three months is a lifetime — and competitors weren’t standing still.

Earlier today, Google released Gemini 3.1 Proan update that brings a key innovation to the company’s horsepower model: three levels of adaptive thinking that effectively turns it into a lightweight version of Google’s special deep thinking system.

The release marks the first time Google has issued a “single point” update to a Gemini model, marking a shift in the company’s release strategy from occasional full version launches to regular updates. Even more important for business AI teams evaluating their model stack, the new 3.1 Pro three-level thinking system – low, medium, and high – gives developers and IT leaders a single model that can measure their dynamic thinking effort, from quick answers to common questions to deep multi-minute brainstorming sessions on complex problems.

The model is now out in preview across the Gemini API with Google AI StudioGemini CLI, Google’s agent development platform Antigravity, Vertex AI, Gemini Enterprise, Android Studio, Gemini consumer apps, and NotebookLM.

The ‘Deep Think Mini’ effect: transformative thinking on demand

The most important feature in Gemini 3.1 Pro is not a single benchmark number – the introduction of a three-tier logic level system that gives users fine control over how much computational effort the model puts into each response.

Gemini 3 Pro offered only two modes of thinking: low and high. The new 3.1 Pro adds a mid setting (similar to the previous height) and, seriously, fixes what “height” means. When set to high, the 3.1 Pro behaves like a “miniature version of Gemini Deep Think” – the company’s special thinking model updated last week.

The impact of business distribution can be significant. Rather than forwarding requests to different specialized models based on the complexity of the task – a common but functionally burdensome pattern – organizations can now use a single storage model and adjust the depth of thinking based on the task at hand. Round-the-clock document summarization can be initiated from low-level thinking with fast response times, while complex analytical tasks can be elevated to high-level thinking with Deep Think-caliber thinking.

Benchmark Performance: More than 3 times more than 3 Pro

The benchmarks published by Google tell the story of impressive improvements, especially in areas related to cognitive ability and agency.

Opened ARC-AGI-2benchmark that tests the model’s ability to solve abstract thought patterns, 3.1 Pro points 77.1% — more than double the 31.1% achieved by the Gemini 3 Pro and well ahead of Anthropic’s Sonnet 4.6 (58.3%) and Opus 4.6 (68.8%). This result also eclipses OpenAI’s GPT-5.2 (52.9%).

The benefits extend across the board. Opened Humanity’s Final TestIn the rigorous academic benchmark, the 3.1 Pro scored 44.4% without tools, up from the 3 Pro’s 37.5% and ahead of both the Claude Sonnet 4.6 (33.2%) and the Opus 4.6 (40.0%). Opened GPQA Diamondscientific knowledge test, 3.1 Pro reached 94.3%, outperforming all competitors in the list.

Where the results are most relevant to enterprise AI teams is in agent benchmarks – analyzes that measure how well models perform when given tools and multi-step tasks, the type of work that increasingly defines manufacturing AI deployments.

Opened Terminal-Bench 2.0testing agent terminal encoding, the 3.1 Pro scored 68.5% compared to 56.9% for its predecessor. Opened The MCP Atlasbenchmark that measures multi-step workflows using the Model Context Protocol, 3.1 Pro achieved 69.2% — a 15-point improvement over 3 Pro’s 54.1% and nearly 10 points ahead of both Claude and GPT-5.2. And so on BrowseComptesting the power of web searches, the 3.1 Pro scored 85.9%, surpassing the 3 Pro’s 59.2%.

Why Google chose the ‘0.1’ release – and what it shows

The decision to make the version itself is noteworthy. Previous releases of Gemini have followed a pattern of date previews – multiple 2.5 previews, for example, before reaching general availability. Choosing to select this update as 3.1 instead of the other 3 Pro preview suggests that Google views the improvement as big enough to warrant a version upgrade, while the “point one” framework sets the expectation that this is an evolution, not a change.

Google’s blog post says 3.1 Pro builds directly on lessons from the Gemini Deep Think series, incorporating techniques from both earlier and later versions. Benchmarks strongly suggest that reinforcement learning plays an important role in gains, especially in tasks like ARC-AGI-2, coding benchmarks, and agent testing – exactly the domains where RL-based training environments can provide clear reward signals.

The model is being released in preview instead of general availability, Google says it will continue to make improvements in areas such as agent workflow before moving to full GA.

Competitive implications of your business AI stack

For IT decision makers who are evaluating the providers of the frontier model, the release of Gemini 3.1 Pro should not only make them rethink which models to choose but also how to adapt to such a pace of change of their products and services.

The question now is whether this release prompts a response from competitors. The original launch of the Gemini 3 Pro last November started a wave of model releases in both open-weight environments.

With 3.1 Pro regaining benchmark leadership in several critical categories, the pressure is on Anthropic, OpenAI, and the open source community to respond — and in the current state of AI, that response is likely measured in weeks, not months.

Availability

Gemini 3.1 Pro is now available in preview with the Gemini API in Google AI Studio, Gemini CLI, Google Antigravity, and Android Studio for developers. Enterprise customers can access it through Vertex AI and Gemini Enterprise. Customers on the Google AI Pro and Ultra plans can access it through the Gemini app and NotebookLM.