Google launches Gemini 3 with new coding app and record benchmark scores

Google on Tuesday unveiled Gemini 3, its most powerful foundation model to date, now rolling out instantly across the Gemini app and Google’s AI-powered search experience. The release arrives just seven months after Gemini 2.5, underscoring the rapid acceleration in the frontier model race — one that has seen OpenAI ship GPT-5.1 last week and Anthropic introduce Sonnet 4.5 only two months earlier.

Google is also preparing a more advanced research-focused variant called Gemini 3 Deepthink, which will be offered to Google AI Ultra subscribers after it completes additional rounds of safety evaluation.

“With Gemini 3, we’re seeing this massive jump in reasoning,” said Tulsee Doshi, Google’s head of product for Gemini. “It’s responding with a level of depth and nuance that we haven’t seen before.”

Early benchmark results support those claims. On Humanity’s Last Exam, a reasoning-heavy benchmark, Gemini 3 scored 37.4 — the highest score ever recorded — passing GPT-5 Pro’s prior record of 31.64. The model also topped LMArena, a human-driven satisfaction benchmark.

Google noted that the Gemini ecosystem now spans more than 650 million monthly active users, and over 13 million developers have integrated the model into their workflow.

Alongside the model upgrade, Google introduced Google Antigravity, a new Gemini-powered coding interface designed for agentic, multi-pane development. The tool resembles emerging agentic IDEs like Warp and Cursor 2.0, combining a ChatGPT-style conversational window with a terminal and a live browser preview.

“The agent can work with your editor, across your terminal, across your browser to help you build that application in the best way possible,” said DeepMind CTO Koray Kavukcuoglu.