google has launched gemini 3, which the company describes as its “most intelligent model.” the large language model includes improved reasoning and a better ability to infer context and user intent, which google says should reduce the number of prompts required to complete tasks. the release introduces a higher-capability variant called gemini 3 deep think, aimed at complex reasoning workloads and certain benchmark challenges.
google said it shipped gemini 3 at scale across its products on launch day. the model is already live in ai mode in google search, though access is limited to google Artificial Intelligence pro and ultra subscribers at launch. gemini 3 is also available in the gemini app, in ai studio and vertex Artificial Intelligence, and in google’s agentic development platform, antigravity. these placements signal that google intends gemini 3 to be the backbone for both consumer features and developer tools.
the company published benchmark results to showcase the update. gemini 3 pro leads the lmarena leaderboard with a 1501 elo score, ahead of xai’s grok and gemini 2.5 pro. gemini 3 deep think performs even better on certain reasoning tests, including the difficult humanity’s last exam benchmark. the gemini app has been redesigned with a new my stuff folder for chats and documents, an improved shopping experience, and experiments with generative interfaces produced by the model. additionally, the gemini agent can perform multi-step tasks such as managing appointments and reminders, organizing an inbox, and conducting online research, reflecting google’s push to make the model capable of handling end-to-end workflows.
