We’re excited to unveil Gemini 2.5, our smartest AI model to date. This first release, an experimental version of 2.5 Pro, sets a new benchmark by debuting at the top of LMArena, outperforming competitors by a wide margin.

What makes Gemini 2.5 Model different? It’s a thinking model — designed to reason through information before responding. That means better accuracy, deeper understanding, and more intelligent interactions.

In AI, “reasoning” goes beyond simple predictions or classifications. It’s about processing complex information, drawing sound conclusions, interpreting context, and making informed decisions.

We’ve long worked on enhancing AI reasoning using methods like reinforcement learning and chain-of-thought prompting. That journey led us to Gemini 2.0 Flash Thinking, our first step toward integrating deeper reasoning capabilities.

Now, with Gemini 2.5, we’ve leveled up. By combining a significantly refined base model with advanced post-training techniques, we’ve created an AI that can tackle complex problems with greater precision. And this is just the beginning — every future model will build on this thinking-first foundation to power more intuitive, context-aware agents.

Introducing Gemini 2.5 Pro

We’re proud to introduce Gemini 2.5 Pro Experimental — our most sophisticated model designed for tackling the most demanding and complex tasks. It currently leads the LMArena, a benchmark grounded in human preference, by a wide margin. That’s a strong signal of its advanced capabilities and refined, high-quality output.

Gemini 2.5 Pro excels in reasoning, coding, mathematics, and scientific problem-solving, consistently outperforming across industry-standard benchmarks. It’s not just smart — it’s a model that thinks.

You can try 2.5 Pro today in Google AI Studio and through the Gemini app for Gemini Advanced users. It’s also on its way to Vertex AI, enabling broader access and deeper integration into production workflows. And soon, we’ll be rolling out pricing options, including higher rate limits to support scaled, enterprise-level use.

Enhanced reasoning

Gemini 2.5 Pro sets a new standard for advanced reasoning, outperforming across a variety of benchmarks — all without relying on costly test-time techniques like majority voting. Its performance shines on complex tasks, taking the lead in math and science evaluations including GPQA and AIME 2025.

It also makes a standout showing on Humanity’s Last Exam — a challenging dataset crafted by hundreds of experts to reflect the outer edge of human knowledge and critical thinking. Gemini 2.5 Pro achieves a best-in-class score of 18.8% among models without external tool use, reinforcing its status as a top-tier thinking model.

Advanced coding

We’ve made a major leap in coding with Gemini 2.5, taking performance well beyond what 2.0 delivered — and we’re just getting started.

Gemini 2.5 Pro is built to shine where it counts: from designing visually dynamic web apps to powering agentic coding workflows, and handling code refactoring and transformation with ease. On SWE-Bench Verified, the industry benchmark for evaluating autonomous coding agents, 2.5 Pro reaches a standout 63.8% score using a custom agent setup.

And here’s where it gets exciting — with just a one-line prompt, Gemini 2.5 Pro can reason through and generate fully functional video game code, producing an executable from scratch. It’s not just code generation — it’s intelligent creation.

Building on the best of Gemini

Gemini 2.5 continues to evolve the core strengths of the Gemini family — combining native multimodality with an extended context window for deeper understanding and broader application.

Starting today, Gemini 2.5 Pro launches with a 1 million token context window (with 2 million on the way), delivering significant performance gains over previous models. It’s built to grasp large-scale datasets and navigate complex tasks across multiple formats — whether it’s text, audio, images, video, or entire codebases.

Developers and enterprises can explore Gemini 2.5 Pro now in Google AI Studio, while Gemini Advanced users can access it directly from the model selector on both desktop and mobile. And soon, it’s arriving on Vertex AI, expanding access for production-scale deployment.

We’re actively listening — your feedback fuels our progress as we work to unlock even more of Gemini’s potential and make AI more powerful, flexible, and helpful for everyone.