Gemini 3 Pro

What it is

Gemini 3 Pro is the first model of Google's Gemini 3 generation. Google released it in preview on 18 November 2025, in a post credited to Sundar Pichai, Demis Hassabis and Koray Kavukcuoglu ^[1]. It works with a context window of 1 million tokens — the amount of text, images or other input the model can keep in view at one time ^[1][2].

Where it launched

The preview was not limited to one product. It went into the Gemini app for everyone, and into AI Mode in Search for Google AI Pro and Ultra subscribers ^[1]. Developers could reach it through Google AI Studio, Vertex AI, Gemini Enterprise and Gemini CLI, and through a new platform called Google Antigravity ^[1]. It was also available in third-party tools including Cursor, GitHub, JetBrains and Replit ^[1]. At launch Google said the Gemini app had more than 650 million monthly users, and that 13 million developers had built with its generative models ^[1].

How it was measured

Google reported results on a set of benchmarks that test different skills. On LMArena, a leaderboard where models are ranked by users comparing pairs of answers, Gemini 3 Pro reached 1501 Elo; on WebDev Arena, which compares models on web-development tasks, it reached 1487 Elo ^[1].

On reasoning tests it scored 37.5% on Humanity's Last Exam without tools, 91.9% on GPQA Diamond and a state-of-the-art 23.4% on MathArena Apex ^[1]. For multimodal understanding and factuality it reported 81% on MMMU-Pro, 87.6% on Video-MMMU and 72.1% on SimpleQA Verified ^[1]. On coding and agent tasks it scored 54.2% on Terminal-Bench 2.0 and 76.2% on SWE-bench Verified, and it topped the Vending-Bench 2 leaderboard ^[1]. These are the company's own reported figures; benchmark scores depend on how a test is run, so they are best read as a snapshot of how the model compared with others at release.

What developers got

Pricing through the API was $2 per 1 million input tokens and $12 per 1 million output tokens for prompts of up to 200,000 tokens ^[2]. The model was also free to use in Google AI Studio, subject to rate limits ^[2]. Alongside the model, the API added a "thinking level" parameter, finer controls over media resolution and stricter validation of thought signatures ^[2]. The first two let developers trade speed against depth of processing and choose how much detail an image or video is read at; the third concerns the internal markers a model attaches to its reasoning steps, which tooling has to handle correctly for multi-step agent workflows to work.

In the Gemini line

Gemini 3 Pro was the opening model of its generation, and the preview period was finite: the preview endpoint gemini-3-pro-preview was shut down on 9 March 2026, with developers moved to gemini-3.1-pro-preview ^[3]. That handover shows the rhythm the family settled into — a preview release, a period in which developers build against it, then a move to a newer version. For users, the more consequential change was that a model of this class arrived at the same time in the consumer app, in Search, in cloud and enterprise tooling and in outside code editors, rather than appearing in a lab first and in products later.