Google announced the release of Gemini 2.5, which represents a new range of AI models that focus on improving reasoning capabilities. The new advancement enables AI systems to halt and perform internal processing before generating a response, which represents a major leap forward in AI technology.

Google introduces its most intelligent AI model yet with the launch of Gemini 2.5 Pro Experimental, which represents a new generation of multimodal reasoning systems. The Gemini 2.5 Pro Experimental model will launch on Tuesday through Google AI Studio and the Gemini app, which is accessible to subscribers of the Gemini Advanced AI service that costs $20 per month.

AI Reasoning Models: The Next Big Leap in Artificial Intelligence

The tech industry has been racing to create competing AI models with better reasoning abilities ever since OpenAI launched its initial AI reasoning model, o1, in September 2024. Anthropic, DeepSeek, Google, and xAI have launched their individual AI reasoning models to enter this competitive space. Extra computing power enables these models to conduct thorough fact-checking and problem analysis prior to generating responses.

AI models that utilize reasoning techniques have shown significant enhancements in addressing intricate mathematical problems and programming difficulties. Experts predict that these models will become essential for creating AI agents that function autonomously while requiring minimal human guidance. The advanced algorithms in reasoning models result in higher operational expenses.

Google launched a “thinking” version of Gemini as part of its AI reasoning experiments during December. Gemini 2.5 marks the company’s highest-level specialized approach to rival OpenAI’s “o” series models.

Performance Benchmarks and Capabilities

Google claims that the new Gemini 2.5 Pro model outperforms their older AI systems and achieves better results than some key competitors during various benchmark evaluations. The company specifically developed this model to create visually compelling web applications and agentic coding programs.

The Aider Polyglot assessment measuring code editing ability resulted in a 68.6% score for Gemini 2.5 Pro. The Gemini 2.5 Pro model excelled beyond the top AI models developed by OpenAI and Anthropic, as well as the Chinese AI research lab DeepSeek.

In the SWE-bench Verified software development skills test, Gemini 2.5 Pro scored 63.8%. The model achieved better results than OpenAI’s o3-mini and DeepSeek’s R1 but could not reach Anthropic’s Claude 3.7 Sonnet performance, which earned 70.3%.

In the multimodal Humanity’s Last Exam, which included thousands of crowdsourced questions from mathematics, humanities, and natural science, Gemini 2.5 Pro achieved a score of 18.8% and surpassed most competing flagship models.

Expanding Context Window and Future Updates

Gemini 2.5 Pro’s launch version maintains a 1 million token context window, enabling it to comprehend about 750,000 words per session, which exceeds the total word count of the “Lord of the Rings” book series. Google intends to extend its input capacity by doubling it soon, which will expand the context window to handle 2 million tokens.

Google keeps the Gemini 2.5 Pro API pricing undisclosed after its announcement of advanced features. The company announced that pricing details will be made available in the upcoming weeks.

Google’s new release marks a strong commitment to AI reasoning, which indicates a trajectory toward more precise analytical artificial intelligence systems that handle complex problem-solving. The tech industry’s competitive landscape drives major companies to advance AI technology through the development of a reasoning model, which will become essential for future intelligent systems.