Google announces Gemini 3.1 Pro with twice the reasoning in advanced tests

Google announced this Thursday (19) the launch of Gemini 3.1 Pro, a significant update to the company’s family of artificial intelligence models. The new model stands out for its advances in advanced reasoning, with more than doubled performance in logic tests compared to the previous Gemini 3 Pro. The preview is now available to developers and paid plan subscribers, focusing on complex tasks that require information synthesis and elaborate problem solving.

The release takes place during the internal testing and validation phase of workflows with autonomous agents. Common Usuários access the version through the Gemini application, with increased limits for those who have the Google AI Pro and Ultra subscriptions. Empresas and developers use the tool via the Gemini API, including platforms such as AI Studio, Vertex AI and Android Studio.

The announcement was made by Google DeepMind, which positions the Gemini 3.1 Pro as the company’s most advanced model for challenges that require intelligence beyond simple answers. The update incorporates improvements in multimodal reasoning, covering text, images, audio, video, and full code repositories.

Performance on reasoning benchmarks

Gemini 3.1 Pro achieved 77.1% on ARC-AGI-2, a benchmark that evaluates the ability to solve entirely new logical patterns. Esse result represents more than double the performance recorded by Gemini 3 Pro, which scored 31.1% in the same metric.

The model also outperformed competitors in rigorous evaluations. No Humanity’s Last Exam scored 44.4%, ahead of models from Anthropic and OpenAI. Esses advances reinforce the position of Gemini 3.1 Pro in tasks that demand specialized knowledge and in-depth reasoning.

Improving logical reasoning allows you to deal with complex problems more consistently. The focus is on practical applications where simple retrieval of trained patterns is not enough.

Advances in agentic and search tasks

In APEX-Agents, which measures performance in long-term professional tasks, Gemini 3.1 Pro registered 33.5%. Esse index indicates improved ability to maintain objectives over extended steps.

The model obtained 85.9% in BrowseComp, an evaluation of agentic internet search combined with the use of tools such as Python. Esse result highlights efficiency in autonomous searches and extraction of relevant information.

These gains facilitate the creation of more reliable autonomous agents. Google emphasizes that the model prioritizes workflows that involve multiple steps and tool integration.

Capabilities in systems synthesis and visual creation

Gemini 3.1 Pro demonstrates the ability to simplify complex systems from text prompts. In demos, it generated scalable SVG animations directly from descriptions, producing lightweight visuals compared to traditional videos.

Another application involved configuring a public telemetry stream to create a functional dashboard that visualizes Estação Espacial Internacional’s orbit in real time. The model processed APIs and data to deliver an operational interface.

The model also encoded interactive 3D experiences. One example simulated the flight of a flock of birds that respond to the user’s hand tracking, integrating computer vision and animation.

Creative Coding Examples

In a literary interpretation test, Gemini 3.1 Pro transformed descriptions from the book “The Morro of the Ventos Uivantes” into a modern portfolio interface. AI captured the atmospheric tone of the work to design a contemporary and functional website.

These demonstrations illustrate the reasoning ability to translate abstract concepts into executable code. The model generates visual and interactive structures based on narrative prompts.

The focus on multimodal synthesis expands uses in design, education, and rapid prototyping.

Access and subscription plans

Gemini 3.1 Pro is in preview and integrated into the Gemini app with expanded limits for subscribers of the Google AI Esses plans offer priority access to the smartest model in the range.

In NotebookLM, the version is also restricted to paid users. The tool benefits from expanded capabilities for document synthesis and generation of insights.

Developers access the preview via API on platforms such as AI Studio and Vertex AI. Empresas use the model in corporate environments with Gemini Enterprise.

Availability for developers and companies

Integration into Vertex AI enables enterprise-scale experimentation. Google highlights that the preview serves to refine flows with autonomous agents before broad release.

The model maintains a context window of 1 million tokens, supporting analysis of large volumes of data. Essa feature remains consistent with previous generations of the Gemini 3 series.

The update strengthens the Google ecosystem of tools for AI applied in professional scenarios.