Google unveils Gemini 2.5 Pro AI model with built-in reasoning capabilities

Google has introduced the Gemini 2.5 Pro Experimental, the newest addition to its line of advanced artificial intelligence models, promising improved reasoning capabilities and impressive benchmark results. The AI model, which is part of the Gemini 2.5 family, was announced by Koray Kavukcuoglu, the Chief Technology Officer of Google DeepMind, through a detailed blog post on Wednesday.

Unlike its predecessors in the Gemini 2.0 series, the Gemini 2.5 series features reasoning capabilities integrated directly into the base model, eliminating the need for specialized “Thinking” models such as the Gemini 2.0 Flash Thinking. Kavukcuoglu noted that Google’s enhanced base model underwent rigorous post-training to ensure it could perform sophisticated reasoning tasks without requiring separate “Thinking” designations.

Although specific technical details around the dataset, architecture, and training methodology remain undisclosed, Google revealed that the Gemini 2.5 Pro scored a remarkable 18.8 percent on the notoriously difficult Humanity’s Last Exam, a gold-standard benchmark for testing AI models. Google claims this score is a new state-of-the-art achievement among models without tool use.

The AI model also outperformed its competition across various benchmarks, including OpenAI’s o3-mini, Grok 3 Beta, Claude 3.7 Sonnet, and DeepSeek R1. It particularly excelled in GPQA Diamond, AIME 2024 and 2025, Aider Polyglot, and MMMU, establishing itself as a powerful competitor in the generative AI landscape.

Furthermore, Gemini 2.5 Pro now ranks highest on the LMArena leaderboard, a platform where developers and AI enthusiasts rate models based on performance and usability. It currently holds the top spot, ahead of popular models like Grok 3 preview, GPT 4.5 preview, Gemini 2.0 Flash Thinking, and Gemini 2.0 Pro.

Beyond natural language processing, the Gemini 2.5 Pro has demonstrated significant improvements in coding performance. The model can now create visually appealing web applications and agentic code applications, enhancing its utility for developers. Additionally, it offers native multimodal support and boasts an impressive context window of one million tokens, allowing it to process and comprehend extensive information without interruption.

The Gemini 2.5 Pro model is currently accessible to developers and enterprises via the Google AI Studio. Meanwhile, Gemini Advanced subscribers can experience the model through Gemini’s web client and apps. Google has also announced plans to roll out the model on its Vertex AI platform in the coming weeks, making it more widely available.