Google's new Gemini 2.5 models can "reason through their thoughts before responding"

Google has unveiled Gemini 2.5, its latest set of AI reasoning models designed to enhance the accuracy and depth of AI responses, not long after Gemma 3. This version has enhanced “thinking” power, where the model sequences tasks before offering solutions.

Gemini 2.5 Pro Experimental, the initial model of this sort, is centered on more advanced reasoning abilities. Allowing time to “think” before responding, it attempts to provide more considered and precise responses, particularly to more challenging questions. This approach has made enormous advancements in terms of mathematics and coding.

Here’s how Gemini 2.5 Pro looks against competitors

In experiments, Gemini 2.5 Pro performed superior to previous iterations and leading competitors. Most notably, its 68.6% score in the Aider Polyglot code editing task outscored OpenAI, Anthropic, and DeepSeek models. It scored 63.8% in the SWE-bench Verified program development task, outscoring OpenAI’s o3-mini and DeepSeek’s R1 but lagging behind Anthropic’s Claude 3.7 Sonnet, which scored 70.3%.

One of the most important upgrades in Gemini 2.5 Pro is its ability to process a 1 million token context window, meaning the model can accept longer inputs—up to around 750,000 words. Google hopes to double the capacity to 2 million tokens.

Gemini 2.5 builds on what makes Gemini models great — native multimodality and a long context window. 2.5 Pro ships today with a 1 million token context window (2 million coming soon), with strong performance that improves over previous generations. It can comprehend vast datasets and handle complex problems from different information sources, including text, audio, images, video and even entire code repositories.

Gemini 2.5 Pro Experimental is currently offered in Google’s AI Studio and Gemini app to customers of the $20-a-month Gemini Advanced plan.