Google launches Gemini 3

Background

When Gemini 3 was announced, Google's AI products were already reaching a very large audience. Sundar Pichai reported that AI Overviews in Search had 2 billion monthly users and that the Gemini app had more than 650 million monthly users ^[1]. That installed base shaped how the model was presented: not only as a research result, but as something that would show up inside products people already use.

What Google announced

On 18 November 2025, Google announced Gemini 3, describing it as "a new era of intelligence" and releasing Gemini 3 Pro in preview first ^[1]. The announcement also covered Gemini 3 Deep Think, a second version aimed at harder reasoning problems, which was initially made available to safety testers rather than to the general public ^[1]. Alongside the models, Google launched Antigravity, a platform for agentic software development, on the same day ^[1][2]. "Agentic" here means the platform is built around AI systems that carry out multi-step tasks on a computer rather than only answering questions.

The numbers Google cited

Google reported that Gemini 3 Pro reached the top of LMArena with a score of 1501 Elo ^[1]. LMArena is a leaderboard built from blind, side-by-side comparisons in which users try two anonymous model outputs and pick the better one; the Elo rating compresses those head-to-head results into a single number, so a higher figure means the model won more of those comparisons against stronger opponents.

On Humanity's Last Exam, Gemini 3 Pro scored 37.5% without tools ^[1]. That benchmark is a collection of very difficult, expert-written questions spanning many subjects. "Without tools" means the model answered on its own, without calling a search engine, a calculator or code execution to help — a stricter setting than allowing it to look things up. Benchmarks of this kind are deliberately hard, so absolute percentages tend to look low; the usual reading is the relative standing rather than the raw number.

Gemini 3 Deep Think was reported at 41.0% on Humanity's Last Exam and 45.1% on ARC-AGI-2 ^[1]. ARC-AGI-2 is a test set built from small visual grid puzzles that are easy for people but require the model to work out a rule it has never seen and apply it to new cases, which is why it is used as a probe of general reasoning rather than memorised knowledge. Deep Think's initial audience of safety testers means the variant was not part of the day-one consumer release ^[1].

Where it shipped

Availability mattered as much as the scores. Gemini 3 Pro came to AI Mode in Search on launch day for Google AI Pro and Ultra subscribers ^[1]. AI Mode is the conversational layer of Search, so this put the new model in front of people through the search box itself, gated by subscription tier rather than released as a separate product.

Why it matters

Three things stand out about the launch. First, the announcement led with measured results on shared public benchmarks — a leaderboard position and named test scores — rather than only with demonstrations ^[1]. Second, Deep Think was handled in stages, with a stronger variant opened to safety testers before general release ^[1]. Third, the model, the search integration and a development platform were announced together, so the release covered both the model and the surfaces it runs on: consumer Search, the Gemini app, and a coding environment for agents ^[1][2]. Combined with the user numbers Pichai cited, the framing was of a model change that would be felt across products already in wide use ^[1].