Google Unveils Groundbreaking Reasoning AI Model Gemini 2.5 Pro

Google has introduced a new line of «reasoning» AI models called Gemini 2.5, which pause to think before delivering their responses.

The series kicks off with the experimental version, Gemini 2.5 Pro, a multimodal neural network hailed as the «most intelligent to date.» It is available on the Google AI Studio developer platform and through the Gemini app for Advanced subscription holders at a monthly cost of $20.

«In Gemini 2.5, we have achieved a new level of performance by combining a significantly enhanced core model with improved post-training,» the company stated in its blog.

Google also mentioned that all of its future models will possess reasoning capabilities.

In the Aider Polyglot test, which assesses programming skills, Gemini 2.5 Pro scored 68.6%, outperforming leading solutions from OpenAI, Anthropic, and DeepSeek.

However, in the SWE-bench Verified benchmark, which evaluates software developers’ skills, Google’s solution fell short compared to Claude 3.7 Sonnet, scoring 63.8% against 70.3%.

In the multimodal test called Humanity’s Last Exam, featuring 1,000 questions across mathematics, humanities, and sciences, Gemini 2.5 Pro scored 18.8%, exceeding the performance of its competitors.

Gemini 2.5 Pro boasts a contextual window of 1 million tokens, allowing the model to process approximately 750,000 words at a time, with plans to double this capacity in the future.

«Reasoning» AI requires more time and computational power to generate precise and comprehensive responses. The first «thinking» neural network was showcased by OpenAI when it released o1 in September 2024, prompting other AI developers like DeepSeek, Anthropic, Alibaba, and others to work on their own solutions.

In March, Google announced a range of AI products and features aimed at the healthcare sector.

Previously, the company trained its AI assistant Gemini to analyze and consider users’ search histories to provide more relevant answers.