Google DeepMind releases Gemini 2.5 as its most intelligent AI model yet, with the debut Gemini 2.5 Pro Experimental leading common benchmarks by clear margins. Thinking models analyze information and draw logical conclusions before responding, yielding better accuracy on complex prompts. Native multimodality and long context let the system absorb text, audio, images, video and code repositories without fragmenting intent.

An 18.8% score on Humanity’s Last Exam places 2.5 Pro at the frontier of knowledge reasoning among models without tool use. Math and science gains arrive without costly test-time tricks, while SWE-Bench Verified hits 63.8% for agentic code tasks. A 1 million token window supports large datasets and cross-source problem solving today, with 2 million tokens imminent.

Developers can use the model in Google AI Studio, while Gemini Advanced subscribers select it from the desktop and mobile dropdown. Pricing and higher rate limits for scaled production use will follow soon, and Vertex AI availability is pending. The release trades hype for measurable leaps in logic and code quality.