Google has announced a significant advancement for its flagship AI model with the introduction of Deep Think mode for Gemini 2.5 Pro, representing a major leap in AI reasoning capabilities.
Deep Think leverages cutting-edge research in parallel thinking techniques, allowing the model to explore and evaluate multiple potential solutions simultaneously before providing a response. This approach mirrors how human experts tackle complex problems by considering various angles and hypotheses.
"It uses our latest cutting-edge research in reasoning - including parallel thinking techniques - resulting in incredible performance," explained Demis Hassabis, CEO of Google DeepMind, during the announcement at Google I/O 2025.
The new mode has demonstrated exceptional results across challenging benchmarks. It achieved an impressive score on the 2025 United States of America Mathematical Olympiad (USAMO), widely considered one of the hardest math benchmarks available. Deep Think also leads on LiveCodeBench, a difficult benchmark for competition-level coding, and scored 84% on MMMU, which tests multimodal reasoning across diverse tasks.
Beyond Deep Think, Google has significantly enhanced security across the Gemini 2.5 family. The company implemented advanced safeguards against indirect prompt injection attacks—where malicious instructions are embedded into data an AI model retrieves. According to Google, their new security approach has substantially increased Gemini's protection rate against these attacks during tool use, making the 2.5 series Google's most secure model family to date.
Google is taking a cautious approach with Deep Think's rollout. "Because we're defining the frontier with 2.5 Pro DeepThink, we're taking extra time to conduct more frontier safety evaluations and get further input from safety experts," the company stated. Initially, Deep Think will be available only to trusted testers via the Gemini API to gather feedback before wider release.
The company also announced improvements to Gemini 2.5 Flash, its more efficient model designed for speed and low cost. The updated version uses 20-30% fewer tokens while delivering improved performance across reasoning, multimodality, code, and long-context benchmarks.