Google has unveiled a new family of AI reasoning models called Gemini 2.5, designed to "think" before responding to queries. The company's most intelligent model to date, Gemini 2.5 Pro Experimental, is available to subscribers of Gemini Advanced (a $20/month AI service) through both Google AI Studio and the Gemini app.
This experimental version, the first release in the 2.5 series, has shown notable improvements in both reasoning and coding. Google claims that Gemini 2.5 Pro outperforms its predecessors and rivals on key metrics, particularly in the domains of web apps and coding, marking its latest move in the ongoing AI arms race against DeepSeek, Anthropic and xAI.
Model Capabilities
According to Google, Gemini 2.5 Pro leads the LMArena leaderboard (a metric that measures human preferences) by a wide margin, showcasing its efficiency and competence. The model excels in standard coding, math and science benchmarks, demonstrating strong reasoning and coding abilities.
Gemini 2.5 Pro is at the cutting edge of benchmarks requiring advanced logic; it outperforms competitors on assessments like GPQA and AIME 2025 without relying on costly test‑time techniques such as majority voting.
It also achieves a state‑of‑the‑art 18.8% across models without tool utilisation on Humanity’s Last Exam, a dataset created by hundreds of subject matter experts to capture the frontier of human knowledge and reasoning.
Gemini 2.5 Pro excels at transforming and editing code, as well as in developing aesthetically appealing web apps and agentic code applications. With a customised agent configuration, Gemini 2.5 Pro scores 63.8% on SWE‑Bench Verified, the industry benchmark for agentic code evaluations.
It is based on Gemini models, which offer excellent inherent multimodality and a large context window. 2.5 Pro is available now with a 1 million token context window (2 million coming soon) and improved performance over prior generations. It can understand large datasets and solve complicated problems using a variety of information sources, like text, audio, photos, video and even entire code repositories.
Recent Updates
Gemini 2.5 was released shortly after Google Gemma 3, the latest model in the Gemma family of open‑weight systems and it succeeds Gemma 2, which debuted last year. In a related development, Google has just introduced its native image‑generation capability in Gemini 2.0 Flash, a tool that fuses multimodal input, advanced reasoning and natural language processing (NLP) to produce high‑quality graphics.
Meanwhile, competitors are also advancing their offerings: OpenAI has added image‑generation features in GPT‑4o and DeepSeek released a new version of its general‑purpose AI model, DeepSeek V3‑0324, which currently ranks first in benchmarks among non‑reasoning models.
