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

On Tuesday, Google released Gemini 3, its latest and astir precocious instauration model, which is present instantly disposable done the Gemini app and AI hunt interface.

Coming conscionable 7 months aft the Gemini 2.5 release, the caller exemplary is Google’s astir susceptible LLM yet, and an contiguous contender for the astir susceptible AI instrumentality connected the market. The merchandise besides comes little than a week aft OpenAI released GPT 5.1, and a specified 2 months aft Anthropic released Sonnet 4.5 — a reminder of the blistering gait of frontier exemplary development.

A much research-intensive mentation of the model, called Gemini 3 Deepthink, volition besides beryllium made disposable to Google AI Ultra subscribers successful the coming weeks, erstwhile it passes further rounds of information testing.

“With Gemini 3, we’re seeing this monolithic leap successful reasoning,” said Tulsee Doshi, Google’s caput of merchandise for the Gemini model. “It’s responding with a level of extent and nuance that we haven’t seen before.”

Some of that reasoning powerfulness is already registering connected autarkic benchmarks. With a people of 37.4, the exemplary marked the highest people connected grounds connected the Humanity’s Last Exam benchmark, meant to seizure wide reasoning and expertise. The erstwhile precocious score, held by GPT-5 Pro, was 31.64. Gemini 3 besides topped the leaderboard connected LMArena, a human-led benchmark that measures idiosyncratic satisfaction.

According to Google, the Gemini app presently has much than 650 cardinal monthly progressive users, and 13 cardinal bundle developers person utilized the exemplary arsenic portion of their workflow.

Alongside the basal model, Google besides released a Gemini-powered coding interface called Google Antigravity, allowing for multi-pane agentic coding akin to agentic IDEs similar Warp oregon Cursor 2.0. Specifically, Antigravity combines a ChatGPT-style punctual model with a command-line interface and a browser model that tin amusement the interaction of the changes made by the coding agent.

“The cause tin enactment with your editor, crossed your terminal, crossed your browser to marque definite that it helps you physique that exertion successful the champion mode possible,” said DeepMind CTO Koray Kavukcuoglu.