Google DeepMind Unveils Gemini 3.1 Pro, Achieving Top Benchmark Scores

Google DeepMind has launched Gemini 3.1 Pro, its latest flagship AI model, which has demonstrated strong performance across various benchmarks. The model achieved the highest score in the ARC-AGI-2 test and showed a significant increase in inference capabilities compared to its predecessor.

Digital dashboard showing a 77.1% score on a benchmark test, representing Gemini 3.1 Pro's performance.

Gemini 3.1 Pro recorded a 77.1% score in the ARC-AGI-2 test, more than doubling the performance of the previous 3.0 Pro version. It also scored 98% in ARC-AGI-1. The model's development included contributions from Tsinghua alumnus Yao Shunyu.

Benchmark Performance and Capabilities

Gemini 3.1 Pro features native full-modality input capabilities and supports a context window of up to 1 million tokens. In the Human Last Exam (HLE), the model scored 44.4% without tool assistance. In the ARC-AGI-2 test, its 77.1% score surpassed other models, including Opus 4.6, which had recently topped the charts.

The model also showed advancements in coding and AI agent performance. It achieved 2887 Elo points in LiveCodeBench Pro and scored 68.5% in Terminal-Bench 2.0, outperforming GPT-5.3-Codex. In APEX-Agents, Gemini 3.1 Pro led with 33.5%.

For long-context processing, Gemini 3.1 Pro scored 84.9% in the MRCR v2 128k test. It also supported a 1M Token test, achieving 26.3%, a capability not available in some competing models. Google DeepMind stated that the new model also has a reduced hallucination rate.

Abstract visual of long, stable data pathways, symbolizing Gemini 3.1 Pro's long-context processing and reduced hallucination.

Application Development

Gemini 3.1 Pro's capabilities extend to practical applications, including the generation of code-based animations. It can create SVG animations from text prompts, which are scalable and have small file sizes. The model can also integrate complex systems, such as building real-time aerospace data dashboards by combining public telemetry data streams.

In interactive design, Gemini 3.1 Pro can generate complex 3D flocking effects purely through code, supporting real-time gesture tracking and generative music. For creative programming, the model can transform literary themes into runnable code, such as designing a modern-style personal homepage based on "Wuthering Heights."

Developer's screen displaying AI-generated SVG animations and 3D flocking effects, illustrating Gemini 3.1 Pro's creative applications.

Early Demonstrations

Early tests by Google UX engineer Michael Chang demonstrated Gemini 3.1 Pro's ability to simulate urban planning, generating an 11-second SVG animation from a single-sentence prompt in three minutes. AI expert Simon Willison also used the model to generate an SVG of a pelican. In 3D spatial reasoning, the model generated improved 3D pixelated Pokémon compared to its predecessor and can create interactive animations depicting a seed germinating and growing.

Gemini 3.1 Pro is now available in Gemini and NotebookLM. Developers can access it through Google AI Studio, Antigravity, and Android Studio. Google DeepMind noted that the release of Gemini 3.1 Pro is an ongoing step in its development process, with user feedback influencing iterations.