Google Launches Gemini 2.5 - New Generation of Thinking AI Models

The new generation is led by Gemini 2.5 Pro Experimental, a multimodal AI model with what the company claims is its smartest thinking ability yet. It will be available starting today, March 26, on the Google AI Studio developer platform, as well as in the Gemini app for Gemini Advanced subscribers ($20/month).

Gemini 2.5 has the ability to “pause to think” before giving an answer. (Photo: Google)

Thinking AI – Google's new direction

Google announced that from now on, all of its new AI models will have built-in thinking capabilities.

Since OpenAI introduced the o1, the first thinking AI model, in September 2024, the tech industry has been racing to match or surpass its capabilities. Anthropic, DeepSeek, Google, and xAI all now have thinking AI models that use additional computing power to examine information and analyze problems before coming up with an answer.

Advances in cognitive AI have allowed models to outperform mathematics and programming. Many technologists believe that this will be an important foundation for AI agents – automated systems that can perform tasks without human intervention. However, cognitive AI also consumes more resources, leading to higher operating costs.

Google previously experimented with thinking AI with a special version of Gemini in December 2024. But Gemini 2.5 is the company’s most serious move yet to compete with OpenAI’s “o” series.

Outstanding performance on multiple criteria

The Gemini 2.5 Pro beats many top competitors on a number of tests. (Photo: Google)

Google claims that the Gemini 2.5 Pro not only outperforms its previous AI models, but also beats many top competitors on a number of tests.

In the Aider Polyglot benchmark, which measures the ability to edit programming code, the Gemini 2.5 Pro scored 68.6%, surpassing the top models from OpenAI, Anthropic, and DeepSeek.

However, in the SWE-bench Verified test of software development capabilities, the Gemini 2.5 Pro scored 63.8%, higher than the OpenAI o3-mini and DeepSeek R1, but still lower than Anthropic's Claude 3.7 Sonnet (70.3%).

On Humanity's Last Exam, a multi-disciplinary test that includes thousands of questions across math, social sciences, and natural sciences, the Gemini 2.5 Pro scored 18.8%, higher than most other leading AI models.

Notably, Gemini 2.5 Pro can process 1 million tokens at a time, equivalent to about 750,000 words – longer than the entire Lord of the Rings novel series. Google also revealed that in the near future, this model will support up to 2 million tokens, significantly increasing the ability to analyze and remember long contexts.

Google has not yet disclosed the API pricing for Gemini 2.5 Pro. The company said it will provide more information in the coming weeks.

Khanh Huyen (Source: Tech Crunch)

Source: https://vtcnews.vn/google-ra-mat-gemini-2-5-the-he-mo-hinh-ai-tu-duy-moi-ar933854.html