Alphabet’s Google DeepMind on 1 August introduced Gemini 2.5 “Deep Think”, calling it the company’s most advanced public reasoning model and the first multi-agent system in the Gemini line. The software is immediately available in the Gemini app to customers on the Google AI Ultra plan, a US$250-a-month tier that also includes 30 TB of cloud storage and other premium tools. Deep Think employs extended parallel “thinking time” and new reinforcement-learning techniques, enabling it to outperform rivals on synthetic tests. Google reports scores of 34.8 percent on Humanity’s Last Exam without external tools, 87.6 percent on the LiveCodeBench V6 coding benchmark and 99.2 percent on the AIME 2025 math test, ahead of OpenAI’s o3 and xAI’s Grok 4 in internal comparisons. A larger, slower variant of the model won a gold medal at the 2025 International Mathematical Olympiad—the first time an AI system achieved that distinction. The public release is tuned for everyday use and achieves bronze-level performance on the same exam, according to Google. The gold-medal version is being shared with a small group of mathematicians and academics for research feedback. Google plans to expose Deep Think, with and without tool integrations, through the Gemini API to selected developers in the coming weeks, tightening competition with OpenAI, xAI and Anthropic as tech companies race to commercialise compute-intensive multi-agent AI systems.