Google Brings Gemini’s Deep Think to Its App

Google now lets you try Deep Think in the Gemini app, although it is currently available only to Google AI Ultra subscribers. It is a version of Gemini 2.5 Pro designed to spend more time on difficult problems before answering.

The feature arrives after the first tests presented at Google I/O and new adjustments based on feedback from selected users. Google says this version is faster and more practical for everyday use than the model it used for hours to solve complex math problems.

More time to think and compare solutions

Deep Think does not simply generate an answer and deliver it. The system explores several ideas in parallel, compares different approaches, and can review or combine its proposals before choosing a final answer.

In simple terms, it is like asking several people to solve the same problem, reviewing their methods, and keeping the strongest solution. That process requires more computing time, so responses may take longer and be considerably more detailed.

Google has also applied training techniques so the model can make better use of these reasoning paths. The goal is to make it especially useful when recalling a fact is not enough and you need to plan, program, design, or improve something step by step.

What it is useful for in practice

The company highlights three types of tasks:

• Iterative design and development: it can help build a website or application in stages, adjusting both its appearance and how it works.
• Math and science: it can explore conjectures, analyze scientific literature, and work on problems that require connecting many ideas.
• Advanced programming: it is useful for challenges where problem formulation, algorithmic cost, and the consequences of each decision matter.

In Google’s internal evaluations, Deep Think reached bronze-level performance in the 2025 International Mathematical Olympiad test. It is a variant of the model that reached the gold-medal standard in that competition, but that version needed hours to reason through the problems.

Google also points to top-tier results in LiveCodeBench V6, a competitive programming test, and in Humanity’s Last Exam, which evaluates knowledge across different fields, including science and mathematics. The company does not present these results as a guarantee of success on every task.

How to try it and what its limits are

Google AI Ultra subscribers can activate it by selecting Gemini 2.5 Pro and checking Deep Think in the writing bar. Access includes a fixed number of prompts per day. The feature can use tools such as Google Search and code execution, and generate much longer responses.

Access is not general yet. Google is also preparing tests through the Gemini API for selected developers and businesses over the coming weeks.

In safety testing, Deep Think performed better than Gemini 2.5 Pro on content safety and tone objectivity, but it also rejected harmless requests more often. That balance will matter: a model that reasons better must distinguish between a dangerous request and a perfectly legitimate one.

For you, the main change is concrete: Gemini can dedicate more resources to complex tasks, but it does not replace human review and is not available without limits. The question now is whether that additional time produces genuinely more useful answers outside benchmarks, especially in programming, research, and creative work.