Google has announced the release of its most intelligent artificial intelligence model to date - Gemini 2.5.
The model can "think," meaning it pauses before answering a question. This allows it to analyze information, consider context, draw logical conclusions, and make informed decisions.
Google paid particular attention to improving the model's programming capabilities. Compared to Gemini 2.0, the new version is much better at building Web applications, editing code, and automating software agents.
Below, the company provides an example of how 2.5 Pro can use its reasoning capabilities to create a video game based on a single prompt.
According to the Aider Polyglot evaluation, which measures the ability to edit code, Gemini 2.5 Pro achieved a score of 68.6%, surpassing the leading AI models from OpenAI, Anthropic, and the Chinese lab DeepSeek.
However, in the SWE-bench Verified test, which measures software development skills, Gemini 2.5 Pro scored 63.8%, beating OpenAI's o3-mini and DeepSeek's R1, but losing out to Anthropic's Claude 3.7 Sonnet, which scored 70.3%.
Image by Google
On the Humanity's Last Exam, a multimodal test with thousands of questions in math, humanities, and science, Google's model scored 18.8%, outperforming most competitors.
Image by Google
Gemini 2.5 Pro is available to developers and enterprises in Google AI Studio and the Gemini Advanced app, and will be available in Vertex AI in the coming weeks.
Google has not yet disclosed the cost of API access to Gemini 2.5 Pro, but promises to share details in the coming weeks.
The company encourages testing and feedback to further improve the model's capabilities.
Previously, Google expanded access to the Deep Research feature in the Gemini chatbot for all users. Now it can be used for free in more than 45 languages, including Ukrainian, without a Gemini Advanced subscription.
