Update (31 March): Google has announced that Gemini 2.5 Pro is now also free to use. Last week, only paying Advanced users had access to Google’s latest model.
Gemini 2.5 Pro is now selectable as one of the LLMs within the Gemini app and Google AI Studio.
Original article, 29 March:
Google has announced that it is updating its flagship series of Gemini artificial intelligence models by introducing an experimental version, Gemini 2.5 Pro.
According to Google, Gemini 2.5 Pro is its most intelligent model. It has ‘think’ capabilities. All future Gemini 2.5 models will be so-called think models. These can split tasks into multiple steps and logically going through these steps before providing an answer. The company has indicated that this leads to better performance and higher accuracy.
Koray Kavukcuoglu, chief technology officer of Google DeepMind, explains to SilliconANGLE that an AI system’s reasoning ability goes beyond mere classification and prediction. It also includes analyzing information and drawing logical conclusions, as well as considering context and nuance and making substantiated decisions.
This reasoning capacity was first introduced in the Gemini 2.0 Flash Thinking Experimental AI model, which the company launched in December. The company used techniques such as reinforcement learning and chain-of-thought prompting to develop this model.
With Gemini 2.0 Flash Thinking, users could activate this thinking function by clicking a button while entering a prompt. After that, the model began to reason. It also showed its thought process so that users could see how it had reached its conclusion.
Google no longer adds the Thinking designation
According to the company, thanks to neural training, Gemini 2.5 Pro Experimental has reached a higher level of performance than the basic model. Google considers it the most advanced model for complex tasks. It achieved the highest score on the LMArena ranking based on human preferences.
The model also scored 18.8% on Humanity’s Last Exam, a dataset designed by hundreds of experts focused on human knowledge and reasoning ability. By comparison, OpenAI’s o3-mini scored 14% here and DeepSeek R1 8.6%. Both o3-mini and R1 are also conceptual models that can handle complex reasoning, as can Gemini 2.5 Pro Experimental.
Kavukcuoglu indicated that the company focuses on coding performance and that Gemini 2.5 is a significant leap forward compared to version 2.0, with even more improvements in the pipeline.
To demonstrate the model’s new capabilities, Google researchers asked it to generate an endless runner video game with a dinosaur in HTML, CSS and JavaScript based on just one prompt. The model completed this in one go.
Expansion to 2 million tokens
The experimental Gemini 2.5 Pro has a context window of 1 million tokens, which means it can process very large documents, audio and video files – approximately 1.5 million words. Google has indicated that it wants to expand this window to 2 million tokens.
Thanks to the large context window and high performance, Gemini 2.5 Pro offers a strong basis for AI agents. This allows them to process large data sets and tackle complex problems more effectively. Because AI agents operate and plan autonomously, the model’s improved reasoning ability will significantly improve their ability to understand data and use tools to perform tasks.
Developers and business users can already experiment with Gemini 2.5 Pro in Google AI Studio, and users of Gemini Advanced can select the model directly on desktop and mobile. In the coming weeks, users of Vertex AI, Google’s managed machine learning platform for building and deploying AI, will have access to the model.