Announcing Gemini 2.5: A New Era in AI Thinking

Gemini 2.5 is a breakthrough thinking model designed to address increasingly complex challenges. Our initial 2.5 release, Gemini 2.5 Pro Experimental, sets new standards on various benchmarks and demonstrates exceptional reasoning and coding skills.

Introducing Gemini 2.5: Our Most Advanced Intelligent AI Model

We are proud to launch Gemini 2.5, a trailblazing AI model. The inaugural 2.5 Pro Experimental version ranks at the top of LMArena by a significant margin, showcasing its state-of-the-art capabilities across diverse benchmarks.

These models excel in reasoning by processing thoughts before delivering responses, resulting in superior performance and accuracy.

Advancements in AI Reasoning

In AI, “reasoning” involves more than just classification or prediction. It encompasses the analysis of information, drawing logical conclusions, considering context and nuances, and making informed decisions.

For years, we’ve sought ways to enhance AI intelligence through techniques like reinforcement learning and chain-of-thought prompting. Building upon this, we introduced Gemini 2.0 Flash Thinking.

Now, Gemini 2.5 represents a leap forward by merging an enhanced base model with improved post-training. We’re integrating these thinking capabilities across all our models to tackle more complex problems and support highly capable, context-aware agents.

Introducing Gemini 2.5 Pro

Gemini 2.5 Pro Experimental stands as our most sophisticated model for tackling intricate tasks. It leads the LMArena leaderboard, reflecting a model of exceptional capabilities and high-quality style. 2.5 Pro excels in reasoning and coding, achieving top scores on coding, math, and science benchmarks.

Available now in Google AI Studio and the Gemini app for Advanced users, Gemini 2.5 Pro will soon be on Vertex AI. Pricing details, allowing for higher rate limits in scaled production use, will follow in the coming weeks.

Enhanced Reasoning Abilities

Gemini 2.5 Pro is at the forefront across benchmarks requiring advanced reasoning, like GPQA and AIME 2025, without relying on costly test-time techniques.

It achieves a remarkable 18.8% on Humanity’s Last Exam, a challenging dataset crafted by experts to test the boundaries of human knowledge and reasoning.

Advanced Coding Proficiency

With Gemini 2.5, we’ve significantly improved coding performance over 2.0, with more advancements on the horizon. 2.5 Pro excels in developing visually appealing web apps, agentic code applications, and code transformation. On SWE-Bench Verified, it scores 63.8% with a custom agent setup.

For example, 2.5 Pro can generate executable video game code from a single-line prompt through its reasoning prowess.

Leveraging Gemini’s Strengths

Gemini 2.5 builds upon the strengths of its predecessors, featuring native multimodality and an extensive context window. 2.5 Pro is available now with a 1 million token context window (2 million soon), exhibiting robust performance surpassing previous models. It can process vast datasets and manage complex problems from myriad sources like text, audio, images, and entire code repositories.

Developers and enterprises can start exploring Gemini 2.5 Pro in Google AI Studio, while Gemini Advanced users can access it on both desktop and mobile. It will be available on Vertex AI shortly.

We welcome your feedback to continually refine Gemini’s impressive capabilities, aiming to make AI even more beneficial.