Google unveils Gemini 3 with advanced multimodal reasoning capabilities

Google DeepMind has announced the launch of Gemini 3, initially presenting Gemini 3 Pro in preview version. The model is available from today in the Gemini app, in Google Search for paid subscribers, and in development tools such as AI Studio and Vertex AI. The company has also introduced Gemini 3 Deep Think, a mode that extends the model's reasoning capabilities and is currently under evaluation before its public release.

Gemini 3 Pro surpasses its predecessor, version 2.5 Pro, across all major evaluation metrics. In rankings where users compare models without knowing their identity, Gemini 3 reached first place with 1,501 points. The model demonstrated capability to solve doctoral-level problems in science, achieving 91.9% accuracy on a set of complex scientific questions and establishing new records in advanced mathematical problem-solving.

One of Gemini 3's main improvements lies in its ability to simultaneously process different types of information. The model can analyze text, images, videos, audio, and programming code within the same query while maintaining coherence in its responses. In tests evaluating this visual and video understanding capability, the system achieved results above 80%. Google highlights that Gemini 3 can work with extensive contexts of up to one million information elements, allowing it to analyze long documents, complete videos, or large volumes of code.

For developers, Gemini 3 introduces significant improvements in code generation and its ability to function as an autonomous assistant. The model leads web development rankings and substantially outperforms the previous version in tests measuring systems' ability to fix bugs in existing code. Google has also launched Antigravity, a platform where agents based on Gemini 3 can autonomously plan and execute complex programming tasks, directly accessing the code editor, terminal, and browser.

Beyond software development, Google emphasizes that Gemini 3 can maintain coherence in tasks requiring multiple steps over time. In simulations evaluating long-term planning, the model demonstrated ability to make consistent decisions over extended periods without deviating from established goals. This feature will be applied in the Gemini app to allow the system to execute everyday tasks such as organizing emails or booking services, always under user supervision.

According to Google DeepMind, Gemini 3 has completed the most comprehensive security evaluation process applied to any of its models. The system shows reduction in tendencies to provide excessively compliant responses, greater resistance to manipulation attempts through malicious instructions, and better protection against misuse. The company collaborated with UK regulatory bodies and specialized security firms to conduct independent evaluations before launch.

Key points

• Gemini 3 Pro reaches first place in performance rankings with 1,501 points
• Achieves 91.9% accuracy on doctoral-level scientific problems
• Simultaneously processes text, images, videos, audio, and code while maintaining coherence
• Leads web development rankings and code bug fixing
• Google launches Antigravity, platform for autonomous programming agents
• Deep Think mode extends reasoning capabilities and is under evaluation
• Available in Gemini app, Google Search, and development tools
• Has completed Google's most comprehensive security evaluation process