Gemini 2.5 Deep Think Release

Google DeepMind has integrated Deep Think into the Gemini app for Google AI Ultra subscribers, enabling advanced reasoning for complex problem-solving. This release introduces a version of the model that utilizes parallel thinking and extended inference time to improve performance in highly technical domains such as mathematics, coding, and scientific research.

Parallel Thinking and Extended Inference

Deep Think enhances Gemini's reasoning capabilities by employing parallel thinking techniques, which allow the model to generate, evaluate, and combine multiple ideas simultaneously before finalizing an answer. This process is supported by two primary technical drivers:

  • Extended "Thinking Time": By increasing the inference time, the model can explore more hypotheses and refine its solutions to complex problems.
  • Reinforcement Learning: Google DeepMind developed novel reinforcement learning techniques that encourage the model to utilize these extended reasoning paths, improving its intuitive problem-solving over time.

Performance and Use Cases

Deep Think is designed for tasks requiring iterative development, strategic planning, and step-by-step refinement. Key areas of strength include:

Coding and Algorithmic Development

Deep Think excels at difficult coding problems where problem formulation, time complexity, and the consideration of tradeoffs are critical. It achieves state-of-the-art performance on LiveCodeBench V6, which measures competitive coding performance.

Scientific and Mathematical Discovery

The model is capable of reasoning through complex scientific literature and formulating mathematical conjectures. While the full version of Gemini 2.5 Deep Think achieved a gold-medal standard at the International Mathematical Olympiad (IMO), the version available in the Gemini app is optimized for daily use and reaches Bronze-level performance on the 2025 IMO benchmark based on internal evaluations.

Iterative Design

Deep Think improves both the aesthetics and functionality of web development tasks by building complex components piece by piece.

Beyond these domains, the model also achieves state-of-the-art performance on "Humanity’s Last Exam," a benchmark measuring expertise across science and math.

Safety and Responsibility

Gemini 2.5 Deep Think demonstrates improved content safety and tone-objectivity compared to Gemini 2.5 Pro. However, internal testing indicated a higher tendency to refuse benign requests. Google DeepMind is conducting frontier safety evaluations and implementing mitigations for critical capability levels to address risks associated with increased model complexity.

Availability and Implementation

Deep Think is available to Google AI Ultra subscribers within the Gemini app. Users can activate it by selecting "2.5 Pro" in the model dropdown and toggling the "Deep Think" option in the prompt bar.

Key operational details include:

  • Tool Integration: Deep Think works automatically with Google Search and code execution.
  • Response Length: The model can produce significantly longer responses than standard versions.
  • API Access: A version of Deep Think (with and without tools) will be released to trusted testers via the Gemini API in the coming weeks for developer and enterprise evaluation.

Sources