Google’s next-generation AI model, Gemini 3 Pro, has attracted significant attention as leaks and rumors continue to swirl online. Over the past several weeks, speculation about its launch has only intensified—especially with recent reports suggesting that Gemini 3 Pro might debut alongside Nano Banana 2, Google’s forthcoming AI image model.
Mid-November saw a notable development: the alleged leak of the official Gemini 3 Pro model card, including detailed benchmark results. The document not only compares Gemini 3 Pro to its predecessor, Gemini 2.5 Pro, but also pits Google’s newest contender against leading rivals, such as Anthropic’s Claude Sonnet 4.5 and OpenAI’s recently released GPT-5.1, the current default in ChatGPT.
Benchmark Leaks Hint at Major Progress
The leaked document appears genuine, having surfaced on social media with screenshots, a direct link to its original Google page, and even an archived downloadable copy for anyone interested. Embedded within the model card, performance benchmarks describe Gemini 3 Pro’s advancements across multiple domains. According to the document, Gemini 3 Pro “significantly outperforms Gemini 2.5 Pro across a range of benchmarks requiring enhanced reasoning and multimodal capabilities.” These results, dated November 2025, indicate a rapid pace of progress, especially given the freshness of OpenAI’s GPT-5.1 release.
Performance: Outshining the Competition
Gemini 3 Pro posts exceptionally high scores in a variety of industry benchmarks, as one would expect from a cutting-edge AI model. In most tests, it outpaces competitors—though there are a few close calls. For instance, Gemini 3 Pro and Claude Sonnet 4.5 are neck-and-neck on AIME 2025, a math benchmark with code execution. Claude Sonnet 4.5 narrowly takes the lead in the SWE-Bench Verified (agentic coding) test, with ChatGPT also performing strongly in this category.
However, the new Gemini model truly shines in several areas, dominating its competitors and predecessor in tasks such as Humanity’s Last Exam (academic reasoning), ARC-AGI-2 (visual reasoning), MathArena Apex (math contest challenges), ScreenSpot-Pro (screen understanding), CharXiv Reasoning (interpreting complex charts), OmniDocBench 1.5 (OCR tasks), LiveCodeBench Pro (competitive coding), Vending-Bench 2 (long-horizon agentic tasks), SimpleQA Verified (parametric knowledge), and MRCR v2 (long-context reasoning).
It’s important to note that while these leaked benchmarks are compelling, they aren’t final until Google officially launches Gemini 3 Pro and publishes the finalized model card, which could include updated scores.
What Makes Gemini 3 Pro Different?
The leaked documentation describes Gemini 3 Pro as the next evolution in the Gemini series: a suite of highly capable models designed for advanced, natively multimodal reasoning. Unlike previous iterations, Gemini 3 Pro leverages a “sparse mixture-of-experts (MoE)” transformer architecture, with built-in support for processing text, vision, and audio inputs seamlessly.
Sparse MoE allows the model to activate only the most relevant parameters for each prompt, improving both performance and computational efficiency. Google has utilized an enormous, diverse dataset—including text, code, images, audio, and video—for training, and incorporated reinforcement learning to further boost reasoning abilities. Training occurred on Google’s specialized Tensor Processing Units (TPUs).
Availability
While there’s still no official release date, the leaked model card notes that Gemini 3 Pro will be available through multiple Google platforms and services, including the Gemini App, Google Cloud/Vertex AI, Google AI Studio, the Gemini API, Google AI Mode, Google Antigravity, and other integrated providers.
