Google has released Gemini 3.1 Pro, an updated version of its Gemini 3 Pro AI model. The company describes it as a step forward in core reasoning, intended for complex tasks where straightforward answers fall short.
The model is now available to consumers through the Gemini app and NotebookLM, though access on those platforms is currently limited to users with Google AI Pro and Ultra subscriptions. Developers can access it in preview through the Gemini API, Google AI Studio, Gemini CLI, the agentic development platform Google Antigravity, and Android Studio. Enterprise customers can use it via Vertex AI and Gemini Enterprise.
Google states that Gemini 3.1 Pro builds on the same core intelligence introduced with Gemini 3 Deep Think. The model scored 77.1% on ARC-AGI-2, a benchmark that tests a model’s ability to solve entirely new logic patterns. According to Google, this is more than double the score achieved by Gemini 3 Pro. On GPQA Diamond, a benchmark for scientific knowledge, the model reached 94.3%. On SWE-Bench Verified, which tests agentic coding, it scored 80.6%.
Google says the improved reasoning has practical uses. The company highlights several example applications:
- Generating animated graphics directly from a text prompt
- Building live dashboards that visualize real-time data, such as satellite tracking
- Prototyping interactive 3D visualizations with user controls
- Translating literary themes into website designs
According to a technical model card published by Google DeepMind, Gemini 3.1 Pro accepts text, images, audio, and video as inputs and supports a context window of up to one million tokens. It produces text outputs of up to 64,000 tokens. The model is based on Gemini 3 Pro.
Google DeepMind also published results of its internal safety evaluations. Across most categories, the model performs similarly to or slightly better than Gemini 3 Pro. One category, image-to-text safety, showed a small decline of 0.33 percentage points. The company reviewed flagged content manually and reports that most losses were either false positives or not classified as serious.
Under Google’s Frontier Safety Framework, the model was assessed for risks in five areas: chemical, biological, radiological and nuclear information; cyber threats; harmful manipulation; machine learning research and development; and misalignment. The model did not reach the critical capability level in any of these domains. In the cyber domain, Google notes that Gemini 3.1 Pro shows increased capabilities compared to its predecessor, which had previously reached the alert threshold. However, the company says it remains below the level required to reach the critical capability level, and that mitigations are in place.
Google is releasing Gemini 3.1 Pro in preview rather than as a general release. The company states it wants to validate updates and make further improvements, particularly in agentic workflows, before a full rollout.
Sources: Google Blog, Google DeepMind, 9to5Google
