Google has just unveiled Gemini 2.5 Pro Experimental, its latest major advance in artificial intelligence to date. This model, designed to improve reasoning and contextual understanding, delivers superior performance on benchmarks in mathematics, science, and coding. Available experimentally on Google AI Studio and in the Gemini app for advanced users, it promises unprecedented capabilities.
Key takeaways:
- Gemini 2.5 Pro Experimental improves reasoning and the analysis of complex data.
- The new model excels at creating applications and transforming code.
- It supports up to 1 million tokens, with an extension to 2 million planned.
- Google must contend with increased competition from DeepSeek and OpenAI.
An artificial intelligence that 'thinks'
With Gemini 2.5 Pro Experimental, Google DeepMind introduces a model designed to enhance AI reasoning ability. This artificial intelligence no longer just classifies and predicts outcomes, it analyze information deeply, draw logical conclusions, and take context into accountThis approach is based on techniques like reinforcement learning and the "chain-of-thought prompting" method, which allows the AI to structure its thinking before providing an answer.
Google had already started this shift with Gemini 2.0 and its Flash Thinking model. Today, Gemini 2.5 Pro pushes this approach further, notably by improving its base model and post-training.
Superior performance in coding and reasoning
Gemini 2.5 Pro excels on several benchmark tests. It outperforms other models on tests in mathematics and sciences, including GPQA and AIME 2025. It also achieves a record score of 18.8% on Humanity's Last Exam, a dataset designed to evaluate AI reasoning abilities.
In coding, Gemini 2.5 Pro marks a major evolution compared to Gemini 2.0. It is capable of create complex web applications and autonomous agents. On the SWE-Bench Verified benchmark, a reference for evaluating AIs in programming, it achieves an impressive score of 63.8%. A notable example is its ability to generate executable code for a video game from a single one-line prompt.
An extended context for better understanding
One of Gemini 2.5 Pro's great strengths lies in its context window. It can process up to 1 million tokens at launch, with an expansion to 2 million planned soon. This capability allows it to ingest large amounts of data from varied sources: text, audio, images, video, and even entire code repositories.
This advancement makes Gemini 2.5 Pro particularly useful for developers and companies handling large datasets. It can analyze complex codebases and respond accurately to advanced issues.
Increased competition in the AI market
The launch of Gemini 2.5 Pro comes amid intense competition among artificial intelligence models. Since the deployment of Gemini 2.0 in December, new players have disrupted the landscape. DeepSeek, with its DeepSeek-R1 model, demonstrated that it was possible toachieve high performance with reduced training cost. This open-source model competes with closed solutions like OpenAI's, which is also preparing new reasoning models.
Google also unveiled Gemma 3 this month, an improved version of its small language model, optimized for mobile devices and offering a context window of 128,000 tokens.
With Gemini 2.5 Pro, Google thus reaffirms its ambition to dominate the advanced AI market, while preparing the future with even more powerful models. The model is already available in experimental form on Google AI Studio and in the Gemini app, and will soon be integrated into Vertex AI.
The article "Gemini 2.5 Pro: Google unveils its most advanced AI" was published on the site Abondance.
