Google Gemini 3.1 Pro first impressions: a 'Deep Think Mini' with adjustable reasoning on demand
Summary
Google has released Gemini 3.1 Pro, an incremental update to its frontier model, introducing a three-tier adjustable thinking system (low, medium, high). This feature allows the model to dynamically scale its reasoning effort, effectively acting as a "mini version of Gemini Deep Think" when set to high, offering deep reasoning capabilities previously requiring specialized models. Benchmarks show significant improvements, with 3.1 Pro scoring 77.1% on ARC-AGI-2 (up from 31.1% for 3 Pro), 44.4% on Humanity's Last Exam (up from 37.5%), and 94.3% on GPQA Diamond. Agentic benchmarks also saw substantial gains, including 68.5% on Terminal-Bench 2.0 and 69.2% on MCP Atlas. The model is available in preview via the Gemini API, Google AI Studio, Vertex AI, and other platforms.
Key takeaway
For AI architects and enterprise AI teams evaluating model stacks, Gemini 3.1 Pro's adjustable reasoning capability simplifies deployment by consolidating diverse task complexities into a single model endpoint. This eliminates the operational burden of routing requests to multiple specialized models, allowing for dynamic optimization of response times and reasoning depth. You should experiment with the different thinking levels to match computational effort with task requirements.
Key insights
Gemini 3.1 Pro introduces adjustable reasoning, allowing a single model to dynamically scale computational effort for diverse tasks.
Principles
- Dynamic reasoning optimizes resource allocation.
- Incremental updates can yield significant performance gains.
Method
The model offers low, medium, and high thinking levels, with "high" mode emulating Google's specialized Deep Think reasoning system for complex problem-solving.
In practice
- Use low thinking for quick responses and routine queries.
- Elevate to high thinking for complex analytical tasks requiring deep reasoning.
Topics
- Gemini 3.1 Pro
- Adjustable AI Reasoning
- AI Model Benchmarks
- Agentic AI
- Reinforcement Learning
Best for: CTO, VP of Engineering/Data, AI Architect, AI Engineer, Machine Learning Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.