Google launches Gemini 3 with major improvements

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

Google dropped Gemini 3, its shiniest, brainiest, “please-don’t-make-us-regret-inventing-this” foundation model yet, straight into the Gemini app and the company’s AI-powered search. 

If this feels fast, that’s because it is: Gemini 2.5 debuted just seven months ago, which in AI years is roughly a decade and a half.

The launch also lands amid an industry-wide AI sprint that’s starting to look like a very nerdy arms race. 

OpenAI rolled out GPT 5.1 just six days ago, Anthropic pushed out Sonnet 4.5 two months before that, and now Google’s here waving benchmark charts like a proud parent at a science fair. 

Spoiler: the charts are good.

A turbocharged, research-focused edition, Gemini 3 Deepthink, is waiting in the wings for Google AI Ultra subscribers. 

But Google says the model needs a bit more safety testing before it’s allowed out of the lab, presumably so it doesn’t start improvising its own definition of “Ultra.”

According to the company, the model’s biggest trick is a serious leap in reasoning. 

“It’s responding with a level of depth and nuance that we haven’t seen before,” it said, sounding both impressed and possibly a little wary. 

Benchmarks back it up: Gemini 3 just posted a record 37.4 on Humanity’s Last Exam, a test meant to evaluate general reasoning. 

For context, the previous record holder, GPT-5 Pro, scored 31.64. Google also claims it now tops LMArena, a human-driven ranking of models based on user vibes and satisfaction.

People are definitely showing up for it. 

The Gemini app now boasts more than 650 million monthly users, and some 13 million developers have incorporated the model into their workflow, which means a not-insignificant portion of the world’s code is now being ghostwritten by an algorithm.

Google also launched Antigravity, a Gemini-powered coding interface that feels like a cross between ChatGPT, a command line, and a browser window that politely shows you how badly you broke your app. 

DeepMind CTO Koray Kavukcuoglu says the agent can glide between editor, terminal, and browser to help developers “build that application in the best way possible.”