Gemini 3 Released as Google’s Most Intelligent AI Models Yet, Outperforms GPT-5.1 and Claude 4.5 Sonnet

Google finally released the Gemini 3 family of artificial intelligence (AI) models on Tuesday. The Mountain View-based tech giant called it the company's most intelligent AI model yet, highlighting that it outperforms its predecessor as well as OpenAI's GPT-5.1 in every single major benchmark. The new AI model brings improvements across various aspects, including reasoning, conversations, coding, mathematics, as well as agentic capabilities. The company highlights that Gemini 3 is a major step forward towards creating complex agentic experiences for users.

Gemini 3: Availability and Features

In a blog post, the tech giant announced the release of the new Gemini 3 AI models. The release comes after multiple Google executives on X (formerly known as Twitter) shared cryptic posts about the release. The series includes two models — Gemini 3 Pro and Gemini 3 Deep Think.

The company highlighted that these models are being released at scale. This means users can experience their capabilities across Gemini apps and website, AI Mode in Search, AI Studio and Vertex AI for developers, as well as a new development platform dubbed Google Antigravity.

Reliance Jio has also announced that all of its user base that has availed the free Google AI Pro plan will get access to Gemini 3 Pro as well.

However, not all of these models will be available to everyone. The company is currently rolling out Gemini 3 Pro, the most advanced variant of the series, in preview. It will be the model available across Google's suite of products. However, Gemini 3 Deep Think, the company's enhanced reasoning mode, is only available to safety testers currently. Even after release, this model is reserved for the Google AI Ultra subscribers.

Coming to performance, the company shared Gemini 3 Pro's benchmark evaluations based on internal testing. Based on the data shared by the company, Gemini 3 Pro outperforms its predecessor and OpenAI's GPT-5.1 on every major benchmark. It also outperforms Claude 4.5 Sonnet in all but two benchmarks (AIME 2025 and SWE-Bench for math and coding, respectively).

Gemini 3 Pro also scored 37.5 percent on Humanity's Last Exam (academic reasoning) benchmark without tools. It is considered the most difficult test for AI models, with the previous high score set at 26.5 percent by GPT-5.1. It also scored 31.1 percent on the ARC-AGI-2 benchmark (visual reasoning puzzles), which is considered a benchmark for artificial general intelligence (AGI). GPT-5.1 held the previous record with a score of 17.6 percent, while Gemini 2.5 Pro scored 4.9 percent on it.

The company says Gemini 3 Pro also makes significant advances in coding, with a major focus area being frontend coding. The company says the AI model can now generate fully functional and interactive websites, apps, and scalable vector graphics (SVGs).

Coming to Gemini 3 Deep Think, the tech giant claims that it is even more advanced than 3 Pro, with a claimed performance of 41.0 percent on Humanity's Last Exam and 45.1 percent on ARC-AGI-2. However, it is currently not available to users.

Apart from these, the company also introduced Google Antigravity, a new agentic development platform that lets developers collaborate with AI agents. It is essentially an AI-powered integrated development environment (IDE) where users can access Gemini 3's reasoning, tool use, and agentic coding capabilities. The agents also have a dedicated isolated environment, complete with an editor, terminal, and browser.

This means agents can autonomously plan and execute complex, end-to-end software tasks simultaneously on users' behalf while validating their own code. Antigravity also comes with the company's latest Gemini 2.5 Computer Use model for browser control and Nano Banana, alongside Gemini 3 Pro.