Gemini 3 has landed.
Just seven months after the release of its predecessor, Gemini 2.5, Google released what it claims is its most-advanced foundation model boasting a “massive jump” in artificial intelligence (AI) capabilities.
Release of the new model, now available through Google’s Gemini app and AI search interface, follows OpenAI’s GPT 5.1 by less than a week and two months after Anthropic’s Sonnet 4.5, underscoring the breakneck speed of AI development.
“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.”
The model has already demonstrated its prowess on independent benchmarks, scoring 37.4 on Humanity’s Last Exam, a test designed to measure general reasoning and expertise. The previous record of 31.64 was held by GPT-5 Pro.
“From a benchmark standpoint, this launch is a blow to OpenAI, dominating their recent GPT 5.1 release. However, benchmarks aren’t everything when it comes to user preferences,” Alex Conway, principal software engineer at DataRobot, said in an email. “It’s yet to be seen whether the model’s speaking style, ability to infer user intent, and customizability will win over users of GPT 5 and 5.1. We can expect OpenAI to try to find ways to compete on the benchmarks, whether that means releasing new models in the months to come, or new inference modes with existing models.”
Google DeepMind CEO Demis Hassabis characterized Gemini’s advancement as graduating from “simply reading text and images to reading the room,” highlighting the model’s enhanced ability to understand context and nuance across different media formats.
The improvements extend beyond processing power. Google introduced Gemini Agent, an experimental feature that can handle multi-step tasks across Google’s ecosystem, from managing calendars to organizing Gmail inboxes. The company envisions users asking the agent to research vacation plans based on email information and book rental cars with minimal human intervention.
For developers, Google introduced Antigravity, a new coding platform that showcases vibe coding. In demonstrations, the tool built interactive applications while generating progress reports along the way. The model’s enhanced coding abilities also enable dynamic view, a feature that can generate fully interactive websites with clickable widgets when asked about topics like historical figures. According to Google, the Gemini app currently serves more than 650 million monthly active users, with 13 million software developers incorporating the model into their workflows.
Gemini 3 Pro is available to all users through the Gemini app’s “Thinking” mode; a more research-intensive version, called Gemini 3 Deep Think, will roll out to Google AI Ultra subscribers after additional safety testing.
