Google DeepMind has introduced Gemini 2.5 Pro, a model built to handle more complex reasoning, coding, and multi-step tasks. In its initial rollout, it’s already making waves, claiming the top spot on the LMArena leaderboard, which ranks models based on human preference.

The model excels across several benchmarks in science, mathematics, and code generation. Notably, it scores 63.8% on SWE-Bench Verified using a custom agent setup and performs strongly on GPQA and AIME 2025. On Humanity’s Last Exam, designed to test reasoning at the edges of human knowledge, it reaches 18.8% without using tools—impressive for a model of its kind.

Gemini 2.5 Pro also delivers significant upgrades in coding. It not only writes and edits code more accurately than Gemini 2.0, but it can also generate fully functional applications including games, starting from just a few lines of input.

One of its standout features is a long context window, starting at 1 million tokens and expected to double over time. This allows it to reason over large, diverse inputs like full codebases, text, images, audio, and video.

Gemini 2.5 Pro is now available in Google AI Studio and the Gemini app for Gemini Advanced users. It’s also coming to Vertex AI soon, with broader access and pricing details to follow.