Gemini 3.1 Pro

Gemini 3.1 Pro is Google’s flagship frontier reasoning model engineered for complex problem-solving, long-horizon agentic workflows, and natively multimodal context processing. Built on a distillation of Google DeepMind’s Deep Think architecture, Gemini 3.1 Pro delivers a massive leap in abstract reasoning, multi-file software engineering, and high-precision structured execution.

Model Overview

Gemini 3.1 Pro represents a focused mid-cycle intelligence update to the Gemini 3 architecture, specifically targeted at reducing generation truncation, increasing ARC-AGI pattern recognition, and improving autonomous tool usage.

• Developer: Google DeepMind
• Model Class: Multimodal Frontier Reasoning Engine
• Release Date: February 2026
• Primary Target Use Cases: Autonomous web research, multi-repository code execution, interactive code-based design (SVG/3D canvas), scientific literature synthesis, and agentic tool workflows.

Core Capabilities & Performance Profile

Dynamic Thinking & Abstract Reasoning

• ARC-AGI-2 Intelligence Jump: Achieves a verified score of 77.1% on the ARC-AGI-2 benchmark—more than double the abstract pattern-matching capabilities of Gemini 3 Pro.
• Configurable thinking_level: Integrates granular control over internal chain-of-thought depth with four API levels: low, medium, high, and max.
• No Truncation Generation: Resolves long-form cutoff limits, allowing the generation of complete, massive single-run responses (e.g., full web apps, complex math derivations) up to 64k tokens.

Native Multimodality & Code-Based Output

• Massive Context & Inputs: Natively ingests text, code, high-frame-rate video (up to 10 FPS), native audio streams, and dense visual PDFs across a 1,048,576 token input window.
• Code-Based Vector & Visual Synthesis: Produces crisp, crisp-scalable animated SVGs, Three.js 3D environments, and functional telemetry dashboards rendered directly through code rather than video files.
• Visual Spatial Grounding: Native support for 2D pixel-coordinate pointing, open-vocabulary object intent tracking, and spatial plan generation for robotics/XR interfaces.

Agentic Execution

• Optimized for high autonomy in CLI, browser, and IDE execution platforms (including Google Antigravity, Android Studio, and custom terminal agent loops).
• Dedicated gemini-3.1-pro-preview-customtools endpoint for parallel tool calling, bash integration, and search grounding.

Technical Specifications