Google introduced Gemini 3 Pro, its most advanced AI model. It understands everything and crushes the competition in tests

After months of speculation, Google announced a new family of large language models that has the potential to give OpenAI’s GPT-5 a run for its money. The so-called Gemini 3 era, as the company states in its press release, begins with the Gemini 3 Pro model, which will gradually become available to all users of the Gemini app and other services of the Californian software giant starting today. “It is the best model in the world for multimodal understanding,” Google writes.

It understands everything – from recipes to sports videos

Gemini 3 Pro can process text, images, video, audio, and program code simultaneously. Thanks to a context window for 1 million tokens, it can analyze entire books, hour-long lectures, or extensive documents at once.

In practice, this means interesting usage possibilities. For example, the model can read and translate handwritten family recipes in various languages and create a digital cookbook from them. It can analyze academic papers and create interactive study materials, including interactive flashcards or visualizations using code.

Athletes will appreciate the model’s ability to analyze videos of their performances. For example, Gemini 3 Pro can break down a tennis match recording, identify weaknesses in technique, and suggest a specific training plan. In multimodal tests, it achieved an 81% success rate on MMMU-Pro and 87.6% on Video-MMMU, confirming its ability to understand complex visual content (GPT-5.1 scored 76% and 80.4% in the same tests).

Smartest Model with PhD-Level Results

Gemini 3 Pro performs exceptionally well in independent intelligence tests. On the acclaimed LMArena leaderboard, it achieved a record 1501 Elo points, surpassing all current AI models including GPT, Claude, and the newly introduced Grok 4.1. In the Humanity’s Last Exam test, which assesses knowledge across scientific disciplines, it achieved a 37.5% success rate without using any tools.

Even more impressive are the results in mathematics, where Gemini 3 Pro set a new standard. On the MathArena Apex benchmark, it scored 23.4%, which represents the current world record among all AI models. In the GPQA Diamond test focused on scientific problems, it achieved a respectable 91.9% correct answers.

The model also excels in factual accuracy. In the SimpleQA Verified test, it achieved a 72.1% success rate, which signifies a significant improvement in answer reliability. Google emphasizes that Gemini 3 Pro answers concisely and directly, providing actual information instead of the usual polite phrases.

Special Deep Think Mode for the Most Complex Tasks

Google also introduced Gemini 3 Deep Think, a special mode for solving extremely challenging problems. In this mode, the model achieves even better results – 41% on Humanity’s Last Exam and an astonishing 93.8% on GPQA Diamond.

The most impressive result is 45.1% on the ARC-AGI-2 test, which measures the ability to solve entirely new, unseen problems. For comparison: when the test creators had 400 people solve problems to establish a human baseline, the average individual success rate was 60%. This demonstrates that AI is gradually approaching human capabilities in abstract thinking.

The Deep Think mode will be available for Google AI Ultra subscribers, but only in a few weeks. The company still needs to evaluate risks based on feedback from trusted testers.

Where can you try Gemini 3 Pro?

Starting today, the model will gradually roll out to users of the Gemini app. Google AI Pro and Ultra subscribers will also gain access through the AI Mode in Google Search, which can now generate interactive visualizations and simulations directly based on a query.

Developers can work with Gemini 3 Pro via Google AI Studio, Vertex AI, and the new Google Antigravity platform. The model is also integrated into popular developer tools like Cursor, GitHub, and JetBrains.

Google emphasizes that Gemini 3 Pro has undergone the most extensive security testing of all the company’s models. It shows better resistance to manipulation using the prompt injection method and a reduced risk of misuse for cyberattacks.

What do you think of the new Gemini 3 Pro model?

Source: Google