- Google unveiled its state-of-the-art Gemini 3 Pro AI model delivering impressive performance across all benchmarks.
- In Humanity's Last Exam, Gemini 3 Pro achieved a whopping 37.5% without any tool use.
- In ARC-AGI-2, Gemini 3 Pro scored 31.1%, outranking both GPT-5.1 and Claude Sonnet 4.5.
After over seven months of anticipation, Google has officially released its state-of-the-art Gemini 3 Pro AI model. According to Google, Gemini 3 represents “a suite of highly-capable, natively multimodal, reasoning models.” With the release of Gemini 3, Google claims it’s”taking another big step on the path toward AGI”.
Gemini 3 Pro Outranks All Frontier AI Models
Talking about the architecture, Gemini 3 Pro is a sparse mixture-of-expert (MoE) model, built on the Transformer architecture. On top of that, Google says Gemini 3 Pro was trained solely on Google’s TPUs, which is impressive.
Now, coming to benchmarks, Gemini 3 Pro has hit it out of the park. In the challenging Humanity’s Last Exam, Gemini 3 Pro achieved 37.5% without any tool use. It even outclassed OpenAI’s latest GPT-5.1 model which scored 26.5%. In LMArena, Gemini 3 Pro has taken the first spot with an ELO score of 1,501 points.
Next, in the new ARC-AGI-2 benchmark, Gemini 3 Pro got 31.1%, again beating GPT-5.1 which received 17.6% only. In the SWE-Bench Verified, Gemini 3 Pro got 76.2%, nearly matching GPT-5.1’s 76.3%. However, in this benchmark, Anthropic’s Claude Sonnet 4.5 continues to lead with 77.2%.
Google is also working to bring Gemini 3 Deep Think to Google AI Ultra subscribers which scored 41% on Humanity’s Last Exam and 45.1% in ARC-AGI-2. In terms of agentic coding, Gemini 3 Pro is the leader in WebDev Arena with 1,487 ELO score. It can do long-horizon, high-level planning to perform multi-step, real-world tasks. Gemini Agent is coming to Google AI Ultra subscribers.
Google also introduced a new Antigravity dev platform which is basically an agent-first development environment. It bundles Gemini 3 Pro, Gemini 2.5 Computer Use model, and Nano Banana image generation model. Agents can directly control the editor, terminal, and the browser to plan tasks and execute code.
Gemini 3 Pro is rolling out in the Gemini app for everyone, starting today. Pro and Ultra subscribers can use the new model in AI Mode in Google Search.
