Gemini 3 Pro Benchmark Scores Leaked Before Launch

It's unclear when Google will unveil Gemini 3 Pro, but the next-generation Gemini AI model has appeared in leaks online for several weeks, prompting speculation that a launch might be imminent. Most recently, reports claimed that Google might release Gemini 3 Pro alongside Nano Banana 2, the next-generation AI image model that also appeared in leaks. Fast-forward to mid-November, and the Gemini 3 Pro model card leaked online, complete with benchmark tests that compare Google's new frontier models against its predecessor Gemini 2.5 Pro, and its main rivals, Anthropic's Claude Sonnet 4.5 and OpenAI's newly released GPT-5.1 model, that's now the default in ChatGPT.

GEMINI 3 PRO MODEL CARD LEAKED pic.twitter.com/YDBpFmg2iH

— cheaty (@cheatyyyy) November 18, 2025

The document appears to be genuine, with X user cheaty posting a screenshot of the benchmark, a link to the Google page the Gemini 3 Pro model card was hosted on, and an archived download link that anyone can use to access the model card. The benchmark is included in the document. "Results: Gemini 3 Pro significantly outperforms Gemini 2.5 Pro across a range of benchmarks requiring enhanced reasoning and multimodal capabilities," the document reads, noting the results are from November 2025. Given that OpenAI released GPT-5.1 only a few days ago, the benchmarks must be very recent.

As expected for a new frontier AI model, Google posted high scores for Gemini 3 Pro in various benchmarks. In fact, Gemini 3 Pro comes out on top in most tests, with only a few exceptions. For example, it ties Claude Sonnet 4.5 in AIME 2025 (math, with code execution test), though it outperforms Anthropic's model in the math test alone. Claude 4.5 also beats Gemini 3 Pro in SWE-Bench Verified (agentic coding), but only by a small margin. ChatGPT also does better here than Google's AI model.

Impressively, Gemini 3 Pro crushes rivals in some of these benchmarks, including its predecessor. Gemini 3 Pro scores significantly higher than competitors in Humanity's Last Exam (academic reasoning), ARC-AGI-2 (visual reasoning puzzles), MathArena Apex (challenging math contest problems), ScreenSpot-Pro (screen understanding), CharXiv Reasoning (information synthesis from complex charts), OmniDocBench 1.5 (OCR), LiveCodeBench Pro (competitive coding problems), Vending-Bench 2 (long-horizon agentic tasks), SimpleQA Verified (parametric knowledge), and MRCR v2 (long context performance).

That said, the leaked Gemini 3 Pro benchmarks can't be considered official until Google launches the new model and releases the finalized model card, which can include updated benchmark scores.

The model card doesn't say when Gemini 3 Pro will be available, but notes that "Gemini 3 Pro is the next generation in the Gemini series of models, a suite of highly capable, natively multimodal, reasoning models. Gemini 3 Pro is now Google's most advanced model for complex tasks, and can comprehend vast datasets, challenging problems from different information sources, including text, audio, images, video, and entire code repositories."

Gemini 3 Pro is a "sparse mixture-of-experts (MoE)" transformer-based model, with "native multimodal support for text, vision, and audio inputs." The document explains that sparse MoE refers to the AI's ability to activate only specific model parameters based on the prompt, which leads to performance improvements and computational efficiency. The document also explains that Google used a "large-scale, diverse collection of data," including text, code, images, audio, and video to train the AI. Reinforcement learning techniques were used to improve reasoning. Also, Gemini 3 Pro was trained on Google's Tensor Processing Units (TPU).

The leaked Gemini 3 Pro model card document says that the new model will be available in the Gemini App, Google Cloud/Vertex AI, Google AI Studio, Gemini API, Google AI Mode, Google Antigravity, and from other providers.