This model surpasses GPT-4.5 in various benchmarks by providing enhanced reasoning capabilities, multi-modal understanding, and extended context processing ranges.
As of March 2025, the experimental ‘Gemini 2.5 Pro’ can process various types of input data such as text, images, audio, and video simultaneously, integrating them logically to generate high-dimensional responses. Google defines this model as a ‘thinking’ next-generation AI, moving beyond the simple response structure of previous models.
In particular, Gemini 2.5 Pro strengthens the ‘multi-step reasoning’ capability, which understands complex contexts and generates responses through logical reasoning. It ensures accuracy and efficiency in high-difficulty tasks like mathematics, science, and program code generation.
The model also shows advancement in code generation and transformation abilities, allowing complex code creation executable with a single prompt line, thus enhancing its practical utility in developer-centric applications. Google applied subsequent reinforcement learning techniques to improve AI’s real-time problem-solving skills.
Gemini 2.5 Pro demonstrated its technological superiority in various benchmarks. It outperformed competitors like Claude 3.7 and o3-mini on the LMArena leaderboard, which evaluates human response preferences, taking first place by a significant margin. It achieved 18.8% in ‘Humanity’s Last Exam’ without tools, 84.0% in ‘GPQA Diamond,’ and 86.7% in ‘AIME 2025,’ proving its precise reasoning capabilities.
On the MRCR benchmark, which assesses long document interpretation, it achieved a 91.5% accuracy with 128K tokens, nearly double the performance of GPT-4.5. In MMMU, measuring multimodal understanding, it recorded an 81.7% correct answer rate, showcasing its competitiveness in simultaneously analyzing various forms of information.
In terms of technical specifications, Gemini 2.5 Pro significantly raises the industry standards, supporting a maximum context window of 1 million tokens, offering far greater information processing than existing models. Google plans to expand this range to 2 million tokens soon, with output generation capable of up to 65,000 tokens, providing more detailed and comprehensive responses.
The model is currently available on a web-based platform for Google AI Studio and Gemini Advanced paid subscribers, with plans to extend to enterprise users via Vertex AI. Mobile platform integration is also expected shortly. Detailed commercialization information, including pricing policies and removal of speed limits, will be announced in the coming weeks.
Google aims to refine Gemini 2.5 Pro’s performance through initial user feedback collected via this phased release strategy, planning to integrate it across the company’s entire service ecosystem. While currently experimental, it serves as a prelude to the full-scale commercialization of reasoning-based AI, drawing industry-wide attention.