Google has added a new “Deep Think” mode to Gemini 3 to give the model a major boost in reasoning. The feature is now available to users on the Google AI Ultra plan inside the Gemini app.
Deep Think focuses on tough problems that need multi-step thinking. Google says the model now handles complex math, science and logic tasks better than previous versions. It works by exploring several ideas in parallel before settling on the final answer. This is the same approach Google tested in earlier Deep Think versions that performed well in global math and coding competitions.
The new mode has also shown strong results in benchmark tests. Gemini 3 Deep Think scored 41% on Humanity’s Last Exam without using any external tools. It also hit 45.1% on the ARC-AGI-2 benchmark, which is known for its difficult logic puzzles and code-based tasks. These numbers suggest that the model is now more reliable in structured reasoning.
To use Deep Think, Ultra subscribers can select the option directly from the prompt bar in the Gemini app. The responses may take a little more time, since the model processes multiple possibilities before answering.
From a user’s point of view, this update pushes Gemini closer to being a tool that can actually think through complicated problems, not just summarise or generate quick replies. It should help anyone who works with advanced math, coding, research, or analytical planning.
At the same time, it is still early for this category of AI. Even with better scores, models can make mistakes or produce incomplete logic when the question is vague. But the improvement in reasoning is noticeable, and the update shows Google’s intention to take the lead in this area again.
