Gemini 2.5 is here, and it is an absolute beast

Today, Google has unveiled Gemini 2.5, and it marks a sizeable advancement in AI capabilities. The initial release introduces an experimental version of Gemini 2.5 Pro, positioned as a state-of-the-art model that has already achieved the top spot on the LMArena leaderboard, surpassing previous benchmarks.

Google says that a core focus of Gemini 2.5 is its enhanced reasoning abilities. These models are designed to “think” through problems, analyzing information and drawing logical conclusions before generating responses. This approach significantly improves accuracy and performance, encompassing the ability to incorporate context, nuance, and informed decision-making.

Google’s development of AI reasoning has evolved through methods like reinforcement learning and chain-of-thought prompting. The introduction of Gemini 2.0 Flash Thinking was a step in this direction, and now Gemini 2.5 builds on this foundation by combining an improved base model with refined post-training processes. The goal is to embed these reasoning capabilities across all Gemini models, enabling them to tackle increasingly complex challenges and support sophisticated, context-aware agents.

Gemini 2.5 Pro sets new standards

Gemini 2.5 Pro Experimental stands out as Google’s most advanced model for handling intricate tasks. Its performance on the LMArena leaderboard, which measures human preferences, indicates this latest version of Gemini has exceptional capabilities and stylistic output. Additionally, 2.5 Pro demonstrates strong reasoning and coding proficiencies, with leading scores on common coding, math, and science benchmarks like GPQA and AIME 2025.

Coding performance is also substantially improved. Gemini 2.5 Pro is capable of creating visually compelling web applications and agentic code applications, as well as performing code transformation and editing. For instance, Google says 2.5 Pro can generate executable code for a video game from a single-line prompt, highlighting its advanced reasoning in practical applications.

Building on Gemini’s Strengths

Gemini 2.5 retains core strengths of other Gemini models, including native multimodality and a long context window. Gemini 2.5 Pro also launches with a 1 million token context window, with plans to expand to 2 million tokens. This enables the model to process vast datasets and handle complex problems from various information sources, including text, audio, images, video, and code repositories.

Developers and enterprises can now begin experimenting with Gemini 2.5 Pro in Google AI Studio, and Gemini Advanced users can access it through the model dropdown on desktop and mobile. It will also be available on Vertex AI in the coming weeks.