Google DeepMind recently announced a major advance in the development of its latest generation of artificial intelligence with the introduction of the Gemini 2.5 model. According to the official blog post, it is a "thinking model" that aims to push the boundaries of what AI can do in complex reasoning and problem solving.
Gemini 2.5 is an artificial intelligence model designed to solve increasingly complex problems. The first model in the 2.5 series, Gemini 2.5 Pro Experimental, achieves a significant lead in commonly used tests.
A revolutionary approach to "thinking" in AI. A key aspect of Gemini 2.5 is its ability to analyse information, draw logical conclusions, take context and nuance into account and make informed decisions before it gives an answer. This new architecture is expected to bring a marked improvement in performance and accuracy across a wide range of tasks.
The first model in this new series is Gemini 2.5 Pro Experimental, which, according to the company, achieves significantly better results in common benchmarks and demonstrates strong reasoning and coding capabilities.
The main new features and improvements in Gemini
- Gemini 2.5 Pro Experimental: This model is presented as the most advanced model for complex tasks. In internal tests and public leaderboards such as LMArena it achieves excellent results, a sign of its high quality and capabilities. The model also shows considerable progress in coding, mathematics and science, where it surpasses previous generations in the relevant benchmarks.
- Improved reasoning capabilities: Gemini 2.5 Pro achieves state-of-the-art results in tests focused on advanced reasoning. Without using techniques that increase computational cost, it leads in mathematics and science benchmarks such as GPQA and AIME 2025. Its result in the Humanity's Last Exam test is also remarkable: it achieved the best score among models without the use of external tools, indicating advanced understanding and the ability to think logically.
-
Progress in coding: Google DeepMind highlights the significant progress in the coding capabilities of Gemini 2.5 compared with the previous version 2.0. The model excels at creating complex web applications and agentic code solutions, as well as transforming and editing existing code. In the SWE-Bench Verified benchmark, Gemini 2.5 Pro with a custom agent setup scores 63.8%, which represents a major step forward in autonomous coding.
As an example, the model's ability to generate executable code for a video game from a single text prompt is cited.. - Architecture and capabilities Gemini 2.5 builds on the proven foundations of previous versions, including native multimodality and support for a long context window. Gemini 2.5 Pro currently ships with a context window of 1 million tokens, with plans to expand it to 2 million soon. This allows the model to process and analyse vast amounts of data from various sources efficiently, including text, audio, images and video.
- Availability for developers and users Gemini 2.5 Pro Experimental is already available to developers in Google AI Studio and to users with a Gemini Advanced subscription in the Gemini app. In the coming weeks it is also planned to be made available on the Vertex AI platform. The company has also announced that pricing information for commercial use with higher rate limits will be published soon.
- Future direction Google DeepMind indicates that the "thinking" capability will gradually be integrated into all future Gemini models, which should make it possible to solve even more complex problems and develop more advanced agents with better contextual understanding.
The introduction of Gemini 2.5 represents a significant step forward in artificial intelligence, particularly in reasoning and solving complex tasks. The results achieved in various benchmarks and the capabilities presented point to the potential for further innovation and applications of this technology.
Source: Gemini 2.5: Our most intelligent AI model. Google Blog
This article summarises the key information from Google DeepMind's official blog post about the Gemini 2.5 model. For more details and the latest information, we recommend visiting the original source.