After waiting for over seven months, Google has finally released its new AI model called Gemini 3 Pro. Google says it is a very powerful AI that can handle many types of information and solve complex problems. They also claim this release is a big step toward building truly general AI.
Gemini 3 Pro is built using something called a mixture-of-expert model on top of the Transformer architecture. Google trained it entirely on their own hardware, called TPUs, which is a big deal.
When it comes to performance, Gemini 3 Pro is doing extremely well. In a tough test called Humanity’s Last Exam, it scored 37.5% without using any extra tools. That is much higher than OpenAI’s latest GPT-5.1, which scored 26.5%. In another benchmark, LMArena, Gemini 3 Pro came first with a score of 1,501 points.
In the ARC-AGI-2 benchmark, Gemini 3 Pro scored 31.1%, beating GPT-5.1 which scored 17.6%. In SWE-Bench Verified, it scored 76.2%, just a little below GPT-5.1’s 76.3%. Anthropic’s Claude Sonnet 4.5 is still ahead here with 77.2%.
Google is also working on a more advanced version called Gemini 3 Deep Think for AI Ultra subscribers. It scored 41% on Humanity’s Last Exam and 45.1% on ARC-AGI-2. Gemini 3 Pro also leads in coding tasks, scoring 1,487 in WebDev Arena. It can plan long tasks and complete multi-step real-world jobs.
Google also launched a new platform called Antigravity. It is made for developers and combines Gemini 3 Pro with other AI models, letting AI agents control the editor, terminal, and browser to plan and run tasks.
Gemini 3 Pro is available now in the Gemini app for everyone. Pro and Ultra subscribers can also use it in AI Mode on Google Search.
