Last week, Google DeepMind announced the release of Gemini 2.5, described as its smartest AI model to date. The first release in this series is Gemini 2.5 Pro Experimental, a model that has already topped the LMArena rankings by a significant margin. This benchmark measures human preferences, indicating that the model not only performs well technically, but also delivers answers in a high-quality style that users prefer.
What sets Gemini 2.5 apart is its classification as a “thinking model,” designed to reason before generating answers. This approach results in improved performance and higher accuracy when tackling complex problems. The model demonstrates exceptional capabilities on challenging tasks, topping math and science benchmarks such as GPQA and AIME 2025, and scoring 18.8% on Humanity’s Last Exam, a dataset created by experts to test the frontiers of human knowledge.
Gemini 2.5 enters an increasingly competitive landscape of AI models focused on reasoning. Anthropic’s Claude 3.7 Sonnet, released in February 2025, incorporates a dedicated reasoning mode that enables extended deliberation on complex questions. OpenAI’s GPT-4.5, released in early 2025, features similar capabilities, emphasizing step-by-step problem solving. Most notably, DeepSeek’s R1 model, released in late 2024, pioneered the commercial implementation of a recursive reasoning framework that allows the model to iteratively refine its own thinking, demonstrating particularly strong performance on mathematical reasoning tasks. These developments collectively represent a significant shift in the AI industry toward models that can more explicitly demonstrate their reasoning processes. In addition to its reasoning capabilities, Gemini 2.5 retains the native multimodality and extensive context window that characterized previous versions. With a context window of 1 million tokens (with plans to expand to 2 million), the model can process and understand vast datasets from multiple sources, including text, audio, images, video, and entire code repositories. This multimodal foundation enables it to tackle complex problems and support more capable, context-aware agents.
Gemini 2.5 Pro is currently available in Google AI Studio and the Gemini app for Gemini Advanced users, with plans to bring it to Vertex AI soon. Google will introduce pricing in the coming weeks to enable production use at scale with higher rate limits.
