Gemini 3 Pro: Search Launch, Multimodal Tools & Scores

One preview reached products across Google

Google released Gemini 3 Pro in preview on November 18. It began rolling out that day in the Gemini app, AI Mode in Search, the Gemini API and AI Studio, Vertex AI, Gemini CLI and a new development environment called Antigravity. AI Mode access in Search was limited to Google AI Pro and Ultra subscribers; this was not a replacement for every ordinary search result. [2]

Atlas interpretation: The launch mattered as a distribution event as much as a model release. One model family appeared in a consumer assistant, search, developer tools and enterprise infrastructure at once. Google could observe different kinds of work without waiting for separate product launches to catch up. [2]

Multimodal meant more than attaching an image

Gemini 3 Pro accepted text, images, video, audio and code with a one-million-token context window. Google showed those inputs feeding outputs such as an interactive guide built from papers and lectures, video-based sports analysis, and interfaces generated inside AI Mode. These were product demonstrations, not measured success rates. [2]

Atlas interpretation: The common thread is translation between forms. The model could turn a long, mixed set of source material into code, an interface or a plan. That is a more useful description than saying it could see and hear, because the result still had to be checked against the source material and the requested task. [2]

The benchmark table was not one race

Google reported 37.5% on Humanity's Last Exam without tools, 81% on MMMU-Pro and 76.2% on SWE-bench Verified for Gemini 3 Pro. Its methodology says Gemini results were generally pass@1 with default API sampling, but competing scores could come from provider reports, public leaderboards or Google's own reruns. SWE-bench systems also used different scaffolds and infrastructure. [2][3]

The launch also introduced Gemini 3 Deep Think to safety testers, but did not yet release it to Ultra subscribers. Its higher scores and code-enabled ARC-AGI-2 result belonged to that separate mode and condition. [2]

Atlas interpretation: A score can establish performance on a named task under a named setup. Aggregating different tool rules, harnesses and reporting methods into a single frontier ranking erases the conditions that make the comparison meaningful. [3]