Google Launches Gemini 2.5 Deep Think Reasoning Model

Google released Gemini 2.5 Deep Think on August 1, 2025, a multi-agent AI reasoning model designed for complex problem-solving in mathematics, coding, and science. The model is available to Google AI Ultra subscribers for $250 per month via the Gemini app, with plans to extend access to developers through the Gemini API in the coming weeks.

Deep Think utilizes parallel thinking techniques, spawning multiple agents to explore various solutions simultaneously before selecting the best answer. Google launched two versions of the technology: a consumer version optimized for speed that achieves bronze-level performance on the International Mathematical Olympiad (IMO) benchmark, and a high-performance variant that achieved a gold medal at the IMO. The latter is provided to a select group of mathematicians and academics for research purposes.

Google claims the model outperforms competitors such as OpenAI's o3 and xAI's Grok 4 on benchmarks like LiveCodeBench 6 and Humanity’s Last Exam, though some reports indicate Grok 4 Heavy scored higher on the latter. The company noted that Deep Think has a higher tendency to refuse benign requests compared to Gemini 2.5 Pro. To support this and other AI growth, parent company Alphabet Inc. increased its projected 2025 capital expenditures to $85 billion for AI infrastructure.