Google Releases Gemini 3.1 Pro: Next-Gen Multimodal Reasoning Model

Google DeepMind has released the Gemini 3.1 Pro model card, the latest iteration in the Gemini 3 series and Google’s most advanced model for complex tasks. As a natively multimodal reasoning model, Gemini 3.1 Pro can process massively multimodal information from text, audio, images, video, and entire code repositories.

Core Specifications

• Context Window: 1 million tokens
• Output Length: 64K tokens
• Input Types: Text, images, audio, video
• Base Architecture: Based on Gemini 3 Pro

Performance

Gemini 3.1 Pro significantly outperforms Gemini 3 Pro and competitors across key benchmarks:

Academic Reasoning

• Humanity’s Last Exam: 44.4% (no tools) / 51.4% (search+code) - Industry leading
• ARC-AGI-2: 77.1% - Abstract reasoning puzzles
• GPQA Diamond: 94.3% - Scientific knowledge Q&A

Coding Capabilities

• SWE-Bench Verified: 80.6% - Agentic coding
• Terminal-Bench 2.0: 68.5% - Terminal code agent
• LiveCodeBench Pro: Elo 2887 - Competitive programming

Multimodal & Long Context

• MMMU-Pro: 80.5% - Multimodal understanding
• MRCR v2 (128k): 84.9% - Long context retrieval
• MMMLU: 92.6% - Multilingual Q&A

Agentic Capabilities

• APEX-Agents: 33.5% - Long horizon professional tasks
• BrowseComp: 85.9% - Agentic search
• MCP Atlas: 69.2% - Multi-step workflows

Use Cases

Gemini 3.1 Pro is particularly well-suited for:

• Agentic performance tasks
• Advanced coding and code generation
• Long context and multimodal understanding
• Algorithm development
• Complex problem solving and strategic planning

Read the full model card at https://deepmind.google/models/model-cards/gemini-3-1-pro/