Google announced on Tuesday the release of Gemini 3, the latest version of its flagship large language model, which it calls its smartest model yet, with a whole lot of benchmarks to back it up. The company said that Gemini 3 Pro will be available to users in preview across several Google products starting today, including being deployed in Search, and the “enhanced reasoning model” Gemini 3 Deep Think will start rolling out to Google AI Ultra subscribers after it completes safety testing.
If you follow any AI-obsessed types on social media, you know Gemini 3 has been wildly anticipated by that bunch, who kept whispering about how the model would be a game-changer. Time will tell if it actually achieves that, but Google has a lot of metrics to show how the model is better on paper than previous iterations. The company bragged that it now tops the leaderboard of LMArena, a benchmarking tool used to compare LLMs, dethroning Grok 4.1 Thinking. Google also claims that Gemini 3 Pro demonstrates “PhD-level reasoning” on Humanity’s Last Exam and GPQA Diamond, and set a new record for mathematics performance on MathArena Apex. The scores, of course, climb even higher with Gemini 3 Deep Think.
Does it matter that AI benchmarks are considered unreliable and misleading? Not if you’re on the hype train, it doesn’t.
So Gemini 3 is smarter, which tends to be the primary upgrade that new LLM models bring. But Google is introducing some new capabilities with this model, too. The company announced Google Antigravity, an integrated development environment (IDE) for coding that has a built-in AI agent for maximizing your vibe coding. Basically, it’s a way to more easily hand off coding tasks to AI, which can work across the editor, terminal, and browser based on the user’s commands. While Gemini 3 Pro will be the default, the company said Antigravity also supports Claude Sonnet 4.5 and GPT-OSS agents. Antigravity will be available for testing starting today on Windows, Mac, and Linux.
In addition to the newly launched Antigravity, developers will be able to access Gemini 3 Pro in Google’s Vertex AI and AI Studio. The company also claims Gemini 3’s coding benchmarks are world beaters, too—topping the WebDev Arena leaderboard and setting a new high on Terminal-Bench 2.0. Again, your mileage may vary as to whether those mean anything to you or if they’re just numbers on a test.
The big thing about Gemini 3 is that you’re just not going to be able to avoid it. For the first time, the company is immediately integrating the model into Search. So any time you get an answer from AI Mode, it’ll have run through Gemini 3.
© GoogleThe Gemini app is where the model will primarily live, of course, and users will find what Google calls a “generative interface,” which offers two output modes called visual layout and dynamic view. The visual layout is a more traditional experience, pulling in images to the output based on the user’s prompt. The dynamic view sees Gemini quickly whip up a website-style interface with functioning buttons and let you interact with different pages of information.
Google’s Gemini 3 is the latest model to join the game of oneupsmanship from the biggest players in the AI space. Earlier this month, OpenAI dropped its GPT-5.1 model, and on Monday, Elon Musk’s xAI released Grok 4.1. It took all of 24 hours before Google grabbed the spotlight. We’ll see how long the company can hold onto it before it’s usurped by another model.
English (US) ·