Google officially launched Gemini 3.1 Pro, marking its most advanced core intelligence update to date. This “.1” release is a significant departure from Google’s traditional half-step updates (like 1.5 or 2.5), signaling a focused breakthrough in reasoning and agentic reliability rather than just a feature expansion.
The model is now rolling out globally across the Gemini app, NotebookLM, and developer platforms.
Key Upgrades: Reasoning & Agentic Power
The headline for Gemini 3.1 Pro is its performance on ARC-AGI-2, a rigorous benchmark for abstract reasoning on entirely new logic patterns.
- Reasoning Leap: It achieved a verified score of 77.1% on ARC-AGI-2—more than doubling the performance of Gemini 3 Pro (31.1%).
- Deep Think Integration: It incorporates the “upgraded core intelligence” first previewed in Gemini 3 Deep Think last week, making it highly capable of solving complex scientific and engineering problems.
- Agentic Coding: In SWE-Bench Verified, it scored 80.6%, placing it as a top-tier model for autonomous software engineering tasks.
New Multimodal Capabilities
| Feature | Description |
| Code-Based Animations | Generates website-ready, animated SVGs directly from text prompts. These are built in pure code, remaining crisp at any scale with tiny file sizes. |
| Interactive 3D Simulation | Can code immersive experiences, such as 3D starling murmurations that respond to hand-tracking and dynamic generative audio. |
| Creative Translation | Capable of translating literary themes into functional code (e.g., designing a modern UI based on the “vibe” of a classic novel). |
| Long Context | Retains the 1 million token context window, now with significantly improved recall (84.9% on MRCR v2 at 128k). |
Performance Comparison (Benchmarks)
Gemini 3.1 Pro reclaimed the lead in several high-stakes AI leaderboards:
- GPQA Diamond (Scientific Knowledge): 94.3% (Leading most models).
- Humanity’s Last Exam: 44.4% (Highest score without tools).
- LiveCodeBench Pro (Competitive Coding): 2887 Elo (A new SOTA for LLM coding).
- Hallucination Rate: Improved by 38 percentage points compared to the Gemini 3 Pro Preview.
Availability & Access
- Consumers: Available now in the Gemini app and NotebookLM. Higher usage limits are provided for Google AI Pro and Ultra subscribers.+1
- Developers: Accessible in preview via the Gemini API in Google AI Studio, Vertex AI, and the new Google Antigravity agentic development platform.
- Ecosystem: Now integrated into GitHub Copilot (Public Preview) and Android Studio.
