On March 25, 2025, Google released Gemini 2.5 Pro, an experimental model the company described as its “most intelligent AI model, now with thinking.” Google said “Gemini 2.5 models are thinking models, capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy.” Koray Kavukcuoglu, CTO of Google DeepMind, was quoted making that claim in the announcement.
Google reported that Gemini 2.5 Pro “tops the LMArena leaderboard - which measures human preferences - by a significant margin,” and leads on math and science benchmarks including GPQA and AIME 2025. On Humanity’s Last Exam, a broad expert-level test, Google said the model scored 18.8 percent, “state-of-the-art across models without tool use.” It launched through Google AI Studio and the Gemini app for Advanced subscribers, with Vertex AI access to follow.
Gemini 2.5 Pro mattered because it brought Google fully into the reasoning-model race that OpenAI’s o1 and DeepSeek-R1 had defined, pairing a built-in thinking process with Google’s long-context and multimodal strengths. It reset competitive leaderboards and underscored that inference-time reasoning had become table stakes for a frontier model.