On a blog post, Google chief executive Sundar Pichai said: “Nearly two years ago we kicked off the Gemini era, one of our biggest scientific and product endeavors ever undertaken as a company. Since then, it’s been incredible to see how much people love it. AI Overviews now have 2 billion users every month. The Gemini app surpasses 650 million users per month, more than 70% of our Cloud customers use our AI, 13 million developers have built with our generative models, and that is just a snippet of the impact we’re seeing.”
Gemini 3 is introduced as Google’s most intelligent model. Demis Hassabis, who leads Google DeepMind, and Koray Kavukcuoglu, its chief technology officer, said it has what they describe as state-of-the-art reasoning. They said it reads intent more effectively so people need fewer prompts to get what they want. They also said it reads text, images, audio, code and video without barriers.
It arrives with stronger results on technical tests. According to the company, Gemini 3 Pro reached a 1501 Elo score at the LMArena Leaderboard. It recorded 37.5% on what the firm calls Humanity’s Last Exam and 91.9% on GPQA Diamond. The model showed 23.4% on MathArena Apex. On multimodal testing, the published results show 81% on MMMU-Pro and 87.6% on Video-MMMU. The company said this means the model grasps complex material in science and maths with high reliability.
Gemini 3 Deep Think is also announced for advanced reasoning. Early results show 41.0% on Humanity’s Last Exam and 93.8% on GPQA Diamond. It reached 45.1% on ARC-AGI-2 during testing. Google said this mode is in the hands of safety testers first before release to paying users.
How Will People Use It?
Gemini 3 is rolling into AI Mode in Search on the first day of release. Pichai said this marks the first time a main Gemini model runs in Search on day one. The company is bringing it to the Gemini app, to developers in AI Studio and Vertex AI, and through a new development platform named Google Antigravity.
Hassabis and Kavukcuoglu said the model can bring ideas to life. They said it helps people learn from handwritten recipes or academic papers and then turn that knowledge into visuals or tools that match how they want to learn. In a sports example, they said Gemini 3 can analyse a pickleball game video and offer advice.
More from News
- These Are The UK’s Most Wanted Tech Jobs In 2025
- Train Like Kane: The “Statsports Academy” App Helps You To Train Like World’s Best Players
- Autumn Budget 2025: Industry Experts React To The Budget
- How Is OpenAI Venturing Into ECommerce?
- Amazon’s Next Investment Of Up To $50 Billion Will Be On US AI Services
- Tech Companies Reveal What Has Been Slowing Business Down In 2025
- Is AI Changing The Kind Of Jobs The UK Will Have By 2035?
- NATO And Google Sign Massive Deal Worth Millions
Developers are told they can build richer interactive applications. The company said Gemini 3 topped the WebDev Arena leaderboard with 1487 Elo and reached 76.2% on SWE-bench Verified. It also reached 54.2% on Terminal-Bench 2.0 which measures tool use to control a computer. Google said builders can work with it in AI Studio, Vertex AI, Gemini CLI and Antigravity, or on third party platforms such as Cursor and JetBrains.
Google said Gemini 3 can act more on its own within clear guidance. It topped Vending-Bench 2 for longer horizon planning tests where AI runs a simulated vending machine business over a year. The company said this means the model can handle multi-step tasks such as bookings or inbox organisation while the user keeps control.
What About Safety And Tools?
Google said Gemini 3 is its most secure model yet. It has gone through what the firm called the most complete safety evaluations in its history. The company said it has reduced flattery and improved its resistance to prompt attacks and cyber misuse. It said independent assessments came from Apollo, Vaultis and Dreadnode, along with checks from safety groups such as the UK AISI.
Developers are given new API controls. Google DeepMind product lead Shrestha Basu Mallick and Philipp Schmid in developer relations said there is a new “thinking_level” setting to raise or lower reasoning depth. They also announced “media_resolution” for visual detail and token usage. They introduced what they call “Thought Signatures” so the model keeps track of its reasoning across tasks. These are encrypted and needed again in later calls for full quality.
Pricing for grounding with Google Search has changed to US$14 per 1,000 search queries. The team said this supports dynamic agent workflows. They advised builders to define instructions clearly and keep temperature settings steady at 1.0 for best behaviour.
Pichai said the company will continue to improve Gemini 3. He said the rollout is only the start and he wants to see what people create with it.
