Google Unveils Gemini 2.5 Pro: Leading AI Model with Advanced Reasoning

Google has announced Gemini 2.5, described as its most intelligent AI model generation yet. The first release is an experimental version of Gemini 2.5 Pro, which has immediately claimed the top spot on the influential human preference leaderboard LMArena by a significant margin and shows state-of-the-art performance across numerous benchmarks.

What Makes it a 'Thinking Model'?

Google emphasizes that Gemini 2.5 models are "thinking models." Building on techniques like reinforcement learning and chain-of-thought, these models incorporate an enhanced reasoning capability, allowing them to analyze information, draw conclusions, and consider context more deeply before responding. This results in improved accuracy and performance, especially on complex problems. Google plans to integrate these thinking capabilities directly into future models.

State-of-the-Art Performance

Gemini 2.5 Pro Experimental demonstrates leading performance across various demanding benchmarks. Beyond its #1 LMArena ranking, it achieves state-of-the-art results in:

• Math & Science: Leading scores on GPQA and AIME 2025 without costly test-time techniques like majority voting.
• Knowledge & Reasoning: A top score (18.8% without tool use) on Humanity’s Last Exam, a benchmark designed by subject matter experts to test the frontiers of human knowledge.

Leap in Coding Capabilities

The new model represents a significant advancement in coding. Gemini 2.5 Pro excels at agentic code applications, code transformation, editing, and creating visually appealing web apps. It scores an impressive 63.8% on SWE-Bench Verified (using a custom agent setup), an industry standard for evaluating real-world coding tasks. Google even showcased its ability to generate executable code for a video game from a single prompt.

Building on Gemini Foundations

Gemini 2.5 Pro retains the core strengths of the Gemini family, including native multimodality (handling text, audio, images, video, code repositories) and a long context window, shipping today with 1 million tokens (with 2 million coming soon). This allows it to process and reason over vast amounts of information from diverse sources.

How to Access Gemini 2.5 Pro

The experimental version of Gemini 2.5 Pro is available now for developers and enterprises via Google AI Studio and for users of Gemini Advanced within the Gemini app (desktop and mobile). It will be rolled out to Vertex AI in the coming weeks. Pricing details for scaled production use with higher rate limits are also expected soon.

This release marks a significant step in Google's AI development, pushing the boundaries of reasoning and practical capabilities in large language models.

Find more details on the Google AI Blog.