Google drops agentic vision for Gemini 3 flash
Google just rolled out Agentic Vision, a new feature for its Gemini 3 Flash AI model.
This update lets the AI actually "think" through image questions by combining visual reasoning with real Python code—so you get answers backed by evidence, not guesses.
You can try it now in Google AI Studio, Vertex AI, or the Gemini app.
Smarter, sharper image analysis
Agentic Vision turns image processing into an active investigation—think of it as the AI planning, acting, and observing to solve what's in a picture.
It can zoom in on tiny details (like serial numbers), crop images on its own, though rotating currently may require an explicit prompt nudge, and helps reduce hallucination.
Plus, it handles huge files (up to 1 million tokens) and works with images, videos, audio clips—even PDFs.
Outperforms older models in coding tasks
On coding challenges like SWE-bench Verified, Gemini 3 Flash scored 78%, beating the previous Gemini 3 Pro.
Google says more upgrades are coming: expect smarter math skills and even web or reverse image search down the line.