Google has unveiled a new family of reasoning AI models, Gemini 2.5, that stop to “think” before giving answers. The company says the first version, Gemini 2.5 Pro Experimental, outperforms OpenAI, Anthropic, xAI, and DeepSeek in general AI tests that measure comprehension, math, coding, and other abilities.
Google has unveiled a new family of reasoning AI models, Gemini 2.5, that stop to “think” before giving answers. The company says the first version, Gemini 2.5 Pro Experimental, outperforms OpenAI, Anthropic, xAI, and DeepSeek in general AI tests that measure comprehension, math, coding, and other abilities.
In his post on X, Google DeepMind CEO Demis Hassabis called Gemini 2.5 Pro “a stunning state-of-the-art model, #1 on LMArena with a whopping +39 ELO score, with significant improvements in multimodal reasoning, coding, and STEM.”
Gemini 2.5 Pro is an awesome state-of-the-art model, no.1 on LMArena by a whopping +39 ELO points, with significant improvements across the board in multimodal reasoning, coding & STEM. You can try it out now in AI Studio https://t.co/lLpF8ToTVJ & @GeminiApp with Gemini Advanced https://t.co/bgjabz8O1u
— Demis Hassabis (@demishassabis) March 25, 2025
Most notably, Gemini 2.5 Pro Experimental outperformed OpenAI o3 mini and Anthropic Claude 3.7 Sonnet on Humanity's Last Exam (HLE), a newly created test designed to combat saturation, or the problem of industry tests becoming too easy for rapidly evolving AI models. HLE is therefore a relatively more difficult test to master; Gemini 2.5 scored 18.8% compared to 14% for the o3 mini (which only assessed text-based tasks, no images) and 8.9% for the Claude 3.7 Sonnet, ZDNET reports .
The new model, which has already topped the Chatbot Arena leaderboard, also outperformed its competitors in general science, math, and coding tests, though typically by smaller margins, which is to be expected given the speed at which new models are accelerating. Google said that Gemini 2.5 Pro Experimental shows improvements in reasoning, multimodality, and agent-based capabilities, even with “single-line prompting.”
The video below shows how 2.5 Pro uses reasoning capabilities to program a video game based on a single clue.
Gemini 2.5 Pro is available with a one million token pop-up window for Gemini Advanced users via Google AI Studio and the Gemini app, and will also be “coming soon to Vertex AI.” The company added that it will release pricing information in the next few weeks.
Recall that Microsoft added AI-based deep research tools to Copilot — Researcher and Analyst.
