Google’s Gemini 3 Deep Think Debuts: A new era of AI reasoning – But how powerful is it really?

Google today announced the release of Deep Think, its latest and most powerful reasoning mode for its generative AI model Gemini 3. The multimodal AI model will only be available to “Ultra” subscribers of the Gemini App at launch.

Gemini 3 in Deep Think is designed to address complicated problems in mathematics, science, and logic reasoning, among other tasks. In order to do this, it “explores parallel hypotheses and reason deeper with multistep thinking.” In other words, the new AI model will attempt to tackle a complex problem by going down multiple hypothetical paths to reason and find the best answer possible to whatever question is asked.

The performance of Gemini 3 Deep Think mode was impressive on some of the most difficult AI benchmarks in existence today. On the AI benchmark Humanity’s Last Exam, Google claims that Gemini 3 Deep Think attained 41.0% accuracy without access to external tools. On another one of the most difficult test suites known as ARC‑AGI‑2 (code execution enabled), Gemini 3 Deep Think achieved 45.1% accuracy. This is a new performance level that has not been seen before in other large language models that are public.

The new AI model is also capable of handling more complex planning problems where it can help synthesize structured plans and engage in advanced coding workflows. The performance of Gemini 3 Deep Think shows that it is a major upgrade over the previous model, Gemini 2.5 Pro.

Google Gemini 2.5 Pro in Ultra was able to showcase “gold-medal” level performance in competitions like the International Mathematical Olympiad (IMO) and the ICPC World Finals. With that being said, it is no surprise that Deep Think would be able to outperform the original Gemini 2.5 Pro model.

Gemini 3 Deep Think will be available to Gemini App users on Android or iOS, or on the Gemini website for Gemini users with G Suite. If a user is an “Ultra” user of the Gemini App and wants to start using Gemini 3 Deep Think, the user just has to open the Gemini app, select “Deep Think” from the prompt bar, and then select Gemini 3 Pro from the model dropdown.

Gemini 3 Deep Think Overview

Gemini 3 with Deep Think enhances the model’s core competencies in multimodal understanding and reasoning with a new internal thought process. Rather than providing a single-shot response, Deep Think processes multiple streams of reasoning, pursuing various hypotheses in parallel and then converging on the most promising solution. This enables the model to navigate the more convoluted problems that need creative thinking, step-by-step planning, and deductive logic.

In addition to reasoning, Gemini 3 with Deep Think also takes advantage of Gemini 3’s complete multimodal capabilities. Gemini 3 Deep Think is capable of multimodal input which means that the model will be able to process not just text input but also image, code, or other relevant inputs.

Gemini 3 in Deep Think is also better at performing “agentic” capabilities that go beyond answering simple questions and responses. Instead, Deep Think mode is also better at more complicated tasks like “structured planning, coding workflows, scientific questions, and more.”

Google Gemini 3 in Deep Think

For now, Google Gemini 3 in Deep Think is only available for paid “Ultra” subscribers of the Gemini App. Gemini 3 users can download the Gemini App for Android or iOS devices.

Gemini 3 with Deep Think is “experimental” and may be changed in the future. However, there is a chance that Google Gemini 3 with Deep Think could be removed in the future depending on how users respond and how it performs.

Gemini 3 with Deep Think is already showing superhuman-level reasoning with impressive capabilities that can be applied in research, coding, development, scientific problem solving, and data analysis. If you are able to get your hands on the Gemini 3 with Deep Think, you can do so right now through the Gemini App.