Google DeepMind has unveiled Gemini 2.5, its most intelligent AI model to date, marking a significant leap in the AI race. This new version is touted as a major milestone in reasoning and decision-making capabilities.
Revolutionary Thinking Capabilities with Gemini 2.5
According to Koray Kavukcuoglu, CTO of Google DeepMind, Gemini 2.5 models are described as “thinking models.” Unlike previous models, these systems can reason through their thoughts before generating responses, leading to improved performance and accuracy.
The reasoning capability extends beyond basic prediction and classification, allowing the AI to analyze information, deduce logical conclusions, and make informed decisions by considering context and nuance.
DeepMind has been developing AI reasoning techniques, such as reinforcement learning and chain-of-thought prompting, which have contributed to the Gemini 2.5 breakthroughs. Gemini 2.0 Flash Thinking, an earlier iteration, laid the groundwork for these enhancements.
“Now, with Gemini 2.5, we’ve achieved a new level of performance,” says Kavukcuoglu. By combining an enhanced base model with post-training improvements, the model is set to tackle more complex challenges and support context-aware agents in future models.
Gemini 2.5 Secures Top Spot on the LMArena Leaderboard
The experimental Gemini 2.5 Pro model has achieved state-of-the-art performance, securing the top spot on the LMArena leaderboard. This is a key metric for assessing human preferences, demonstrating the Gemini 2.5 Pro’s ability to handle complex tasks with a high level of accuracy and style.
Gemini 2.5 Pro: A Master of Maths, Science, Coding, and Reasoning
Gemini 2.5 Pro has set new benchmarks in the AI race, excelling in mathematics, science, coding, and reasoning. In benchmarks like GPQA and AIME 2025, it leads without using costly techniques like majority voting. It achieved an 18.8% score on Humanity’s Last Exam, which assesses the frontier of human knowledge and reasoning capabilities.
DeepMind has focused on improving coding performance in Gemini 2.5. Compared to its predecessor, Gemini 2.0, this model represents a substantial leap forward in generating visually compelling web applications, transforming code, and coding agents. It even managed to create a video game from a single-line prompt.
Gemini 2.5 Pro excelled on SWE-Bench Verified, the industry standard for agentic code evaluations, scoring 63.8% using a custom agent setup.
Building on Strengths: Multimodality and Long Context Window
Gemini 2.5 builds on the core strengths of earlier Gemini models, including native multimodality and an extended context window. The 2.5 Pro version launches with a one million token context window, with plans to expand to two million tokens soon. This allows Gemini 2.5 to process vast datasets and complex problems from various sources, including text, images, audio, video, and code repositories.
Access and Feedback for Gemini 2.5 Pro
Developers and enterprises can experiment with Gemini 2.5 Pro in Google AI Studio. Additionally, Gemini Advanced users can access it on both desktop and mobile platforms via the model dropdown. Over the coming weeks, it will also be available on Vertex AI.
Google DeepMind encourages feedback from users to refine and enhance Gemini’s capabilities, ensuring it remains at the forefront of the AI race.
