On November 18, 2025, Google introduced Gemini 3, its new flagship family of large multimodal models positioned as its most capable system so far and deployed from day one across Search, the Gemini app, AI Studio, Vertex AI, the Gemini CLI, and the Antigravity IDE. Unlike earlier Gemini releases that appeared in a small set of products first, Gemini 3 arrives as a unified platform that underpins both consumer and enterprise experiences.
Gemini 3 currently centers on Gemini 3 Pro, with Deep Think positioned as a higher-intensity reasoning mode that will roll out to premium and Ultra tiers. Google describes Gemini 3 Pro as its primary model for multimodal understanding and agentic coding, targeting tasks that combine text, code, and rich media. Deep Think is presented as an offline-style mode for the hardest reasoning workloads, including demanding benchmarks and long-horizon planning.
Gemini 3 Deep Think is next level. Deep Think was the engine behind our gold medal-level wins at IMO and ICPC, and now powers an even stronger version of Gemini 3. SOTA above SOTA. – Quoc Le
From an API perspective, Gemini 3 Pro accepts text, images, video, audio, and PDFs within a context window of up to 1,048,576 tokens, with output capped at 65,536 tokens. The same core model is exposed through the Gemini API, Firebase AI Logic, Vertex AI, and Gemini Enterprise, so teams can choose the integration surface that matches their existing infrastructure. The model supports structured JSON outputs and can be combined with built-in tools.
The Gemini 3 Pro model card and related technical overviews highlight state-of-the-art or near state-of-the-art scores on a range of public benchmarks, including exam-style and scientific reasoning tasks. Deep Think pushes many of these numbers further, particularly on long-horizon reasoning tests designed for agents rather than single prompts.
There’s sort of this feeling that Google, which kind of struggled in AI for a couple of years there — they had the launch of Bard and the first versions of Gemini, which had some issues — and I think they were seen as sort of catching up to the state of the art. And now the question is: is this them taking their crown back? — Kevin Roose, Hard Fork
Gemini 3 Pro can analyze combined inputs of text, media, and documents in one request, so developers can send long PDFs, screenshots, and video snippets without building separate pipelines for each modality. This is positioned to unify workloads such as document analysis, log triage, and media-heavy analytics under a single model rather than maintaining separate vision, speech, and language systems.
Gemini 3 Pro is also being integrated into Gemini Code Assist and Gemini CLI. Code Assist users in common IDEs are receiving Gemini 3 in agent mode first, with the model responsible for running multi-step coding tasks rather than only inline completions. In the terminal, Gemini CLI exposes the same model for workflows like scaffolding applications, refactoring, documentation generation, and lightweight agents.
Google has highlighted Gemini 3’s ability to plan and execute long-running tasks across tools, including financial analysis, supply-chain planning, and contract review. Benchmarks focused on agents and computer use, such as simulated operations and revenue tasks, are used to show performance in environments where models must interact with user interfaces and external systems.
Developer forums highlight improvements in math-heavy workloads, screen-based tasks, and code-heavy projects, while also debating the risk of benchmark contamination and the gap between synthetic evaluations and daily development work. Others note that behavior can be inconsistent and recommend running internal evaluations before committing.
For developers looking to learn more, you may refer to official documentation and model cards.
