gettyimages-2158306662
J Studios/Getty Images
Moments after DeepSeek released its latest model, another AI giant has already stolen back some of the limelight.
On Tuesday, Google announced Gemini 2.5, its "most intelligent" model. The company announced that this initial release is an "experimental version of 2.5 Pro, which is state-of-the-art on a wide range of benchmarks and debuts at #1 on LMArena by a significant margin."
Also: I tried ChatGPT's new Advanced Voice Mode update - here's what changed
A family of thinking models, meaning they reason through their responses, the release follows Google's Gemini 2.0 Flash Thinking, which landed in December.
Most notably, Gemini 2.5 Pro Experimental outperformed OpenAI's o3 mini and Anthropic's Claude 3.7 Sonnet on Humanity's Last Exam (HLE), a recently created benchmark designed to combat saturation, or the problem of industry tests becoming too easy for rapidly evolving models. HLE is, therefore, a relatively harder test to perform well on; Gemini 2.5 scored 18.8% compared to o3 mini's 14% (evaluated using text problems only, no images) and Claude 3.7 Sonnet's 8.9%.
Already topping the Chatbot Arena leaderboard, the new model also outperformed competitors on common benchmarks for science, math, and coding, though usually by a smaller margin, which is now expected given the rate at which new models are accelerating. Google reported that Gemini 2.5 Pro Experimental shows improvements in reasoning, multimodal, and agentic capabilities, even from a "single line prompt."
final-2-5-blog-1-width-1000-format-webp
Google said Gemini 2.5 Pro is available today with a one million token context window for Gemini Advanced users via Google AI Studio and the Gemini app, and will be "coming to Vertex AI soon." The company added that it will release pricing information in the next few weeks.
Want more stories about AI?Sign up for Innovation, our weekly newsletter.
Artificial Intelligence