Google DeepMind has officially launched Gemini 2.5 Deep Think, its most advanced AI reasoning model to date, capable of tackling complex problems by deploying multiple AI agents in parallel. Initially revealed at Google I/O 2025, the system is now available to Ultra Plan subscribers through the Gemini app, with wider access to developers via the Gemini API expected in the coming weeks.
Built for tasks demanding strategic planning and iterative reasoning, Gemini 2.5 Deep Think was used to earn Google a gold medal at the 2025 International Math Olympiad. The company also plans to release the Olympiad-tuned version to select academics for research feedback.
In benchmark testing, Gemini 2.5 Deep Think outperformed rivals from OpenAI, xAI, and Anthropic, including top scores on Humanity’s Last Exam (HLE) and LiveCodeBench 6. It integrates tools like code execution and Google Search, and is designed to produce longer, more detailed outputs than prior models.
The launch marks Google’s push to commercialize multi-agent AI, a high-performance but resource-intensive approach that major AI labs, including xAI and OpenAI, are now also exploring.
