On Tuesday, Google made available its most advanced artificial intelligence (AI) model, Gemini 3, which significantly outperforms OpenAI’s current best AI model, GPT-5.1 Thinking, in several tests.
“Today we are taking another big step at AGI [mesterséges általános intelligencia] with the release of Gemini 3,” wrote Demis Hassabis, CEO of Google DeepMind, and Koray Kavukcuoglu, CTO of DeepMind, in a blog post announcing Gemini 3. “Gemini 3 can bring any idea to life: it quickly understands our intent and the context of our request, so we get what we need with less prompting,” wrote Sundar Pichai, CEO of Alphabet (Google) in the on X.
After receiving the 2024 Nobel Prize in Chemistry from Demis Hassabis, which was awarded to him for the development of the AlphaFold protein research artificial intelligence model
Photo: JONATHAN NACKSTRAND/AFP
Gemini 3 Pro, like other advanced chatbot models, is a so-called big argument model. As we explained in detail in the second episode of the Qubit AI News program, they think longer before answering our questions, which allows them to solve complex mathematical and scientific problems.
As the Verge article points out, the Gemini 3 Pro is an inherently multimodal AI model, meaning it can process text, images and audio simultaneously, instead of handling them separately. According to Google, you can use Gemini 3 Pro to translate recipes, then compile a cookbook from them, or create interactive flashcards from a series of video lectures.
Google’s new AI model is already available with the entry-level subscription for the Gemini chatbot, Google AI Plus, while we can use it even more with the slightly more expensive Google AI Pro, priced similarly to the ChatGPT Plus subscription.
It outperforms OpenAI’s best model in several tests
On the GPQA Diamond test composed of PhD-level scientific questions, the Gemini 3 Pro achieved 91.9 percent, and the Gemini 3 Deep Think version achieved 93.8 percent. The top holder so far was ChatGPT’s GPT-5 Pro model, with 88.1 percent.
Gemini 3 Pro’s capabilities in various tests compared to other AI models
Illustration by Google
On the ARC-AGI-2 test, which measures general intelligence by solving puzzles, Gemini 3 Pro produced a result of 31.1 percent, almost twice as good as OpenAI’s GPT 5.1 Thinking model, which recently achieved 17.6 percent. And Google wasn’t satisfied with that, as the Gemini 3 Deep Think model reached 45.1 percent in the test, which is already close to the average human performance.
Gemini 3 Pro and Gemini 3 Deep Think on the ARC-AGI-2 test compared to other models, including OpenAI’s GPT-5.1-Thinking and GPT-5-Pro models (in blue)
Illustration: ARC Prize
Based on the results the Gemini 3 Pro achieved in these tests, it doesn’t seem like the development of argumentative models has slowed down or stopped yet. Google DeepMind’s Oriol Vinyals wrote on X that the secret to Gemini 3’s increased capability lay in improving pre-training and post-training, and they see room for further growth in both.
“Congratulations to Google on Gemini 3! Looks like a great model,” wrote OpenAI CEO Sam Altman at X on Tuesday in response to Google’s announcement.
