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

On Tuesday, Google released Gemini 3the latest and most advanced foundation model, now immediately available via the Gemini app and AI search interface.

Just seven months later the Gemini 2.5 releasethe new model is Google’s most capable LLM yet, and a direct contender for the most capable AI tool on the market. The release also comes less than a week after OpenAI released GPT 5.1, and just two months after Anthropic released Sonnet 4.5 – a reminder of the blistering pace of frontier model development.

A more research-intensive version of the model, called Gemini 3 Deepthink, will also be made available to Google AI Ultra subscribers in the coming weeks, once it passes further security testing.

“With Gemini 3, we see a huge leap in reasoning,” said Tulsee Doshi, Google’s head of product for the Gemini model. “It responds with a level of depth and nuance that we haven’t seen before.”

Some of that reasoning power is already registered on independent benchmarks. With a score of 37.4, the model achieved the highest score ever the Humanity’s Last Exam benchmarkintended to capture general reasoning and expertise. The previous high score, achieved by GPT-5 Pro, was 31.64. Gemini 3 also topped the leaderboard on LMArena, a human-run benchmark that measures user satisfaction.

According to Google, the Gemini app currently has more than 650 million monthly active users, and 13 million software developers have used the model as part of their workflow.

In addition to the basic model, Google has also released a release a Gemini-powered encryption interface called Google Antigravityallowing multi-window agentic coding, similar to agentic IDEs such as Warp or Cursor 2.0. Specifically, Antigravity combines a ChatGPT-style prompt window with a command-line interface and a browser window that can show the impact of changes made by the coding agent.

“The agent can work with your editor, on your terminal and in your browser to ensure that it helps you build that application in the best possible way,” says DeepMind CTO Koray Kavukcuoglu.