Google unveils Gemini 3 with record-breaking benchmark scores

Google has launched Gemini 3, its latest and most advanced foundation model, making it immediately available through the Gemini app and its AI-powered search interface. The release arrives just seven months after Gemini 2.5 and positions the new model as Google’s most capable large language model to date — and a direct challenger in an increasingly competitive frontier-AI landscape. It also follows closely on the heels of OpenAI’s GPT 5.1, released less than a week ago, and Anthropic’s Sonnet 4.5, launched two months earlier, underscoring the rapid cadence of model development across the industry.

Google will also introduce a more research-intensive variant, Gemini 3 Deepthink, for Google AI Ultra subscribers in the coming weeks, pending additional rounds of safety evaluation.

Tulsee Doshi, Google’s head of product for the Gemini model, informed that Gemini 3 demonstrates a substantial leap in reasoning capability, responding with depth and nuance not previously observed in Google’s systems. That progress is already reflected in independent testing. The model achieved a score of 37.4 on the Humanity’s Last Exam benchmark — the highest score ever recorded, surpassing the previous record of 31.64 held by GPT-5 Pro. Gemini 3 has also taken the top position on LMArena, a human-evaluated benchmark focused on user satisfaction.

Google reports that the Gemini app now reaches more than 650 million monthly active users, while 13 million software developers have incorporated Gemini models into their workflow.

Alongside the flagship model, Google unveiled a new coding interface called Google Antigravity, powered by Gemini and designed to facilitate multi-pane agentic coding similar to advanced IDEs such as Warp or Cursor 2.0. Antigravity combines a conversational prompt window with a command-line terminal and an integrated browser pane capable of displaying real-time changes generated by the coding agent. As per TechCrunch, DeepMind CTO Koray Kavukcuoglu stated that the system is built to operate across the editor, terminal and browser, ensuring the agent can support users in constructing applications as effectively as possible.