Google unveiled Gemini 3, a new generation of multi-models with a focus on context and security

Google is officially launching Gemini 3 — its new and most advanced generation of AI models to replace Gemini 2.5, unveiled at Google I/O in the spring.

Gemini 3 will begin rolling out in AI Mode inside Google Search (for Google AI Pro and AI Ultra subscribers), as well as in the Gemini — app for all users. CEO Sundar Pichai said the model is now «much better at understanding the context and intent of a query, so the user needs less refinement».

Google says Gemini 3 — «the world’s best multimodal AI» providing deeper insights and richer visualizations. First to be released is Gemini 3 Pro in «preview» status. It will be followed by Gemini 3 Deep Think, which is currently being tested by security experts. It will then become available to AI Ultra subscribers.

The company claims the Gemini 3 Pro «significantly outperforms» the Gemini 2.5 Pro on all key AI benchmarks. The model gives more accurate, thoughtful and direct answers, avoiding boilerplate. Deep Think, in turn, performs even better, but is slower due to the increased depth of reasoning.

At the same time, Deep Think performs even better, but is slower due to the increased depth of reasoning.

Gemini 3 was originally designed as a model that can seamlessly combine data from different sources &#8212 text, images, video, audio and code to enhance learning and analytics capabilities.

Google also claims it is the most secure AI in its history. It has passed the «most extensive set of security checks» of any Google model. Gemini 3 demonstrates reduced give-and-take, resistance to prompt attacks and improved protection against abuse, including attempted cyberattacks.