Testing begins for Gemini 2.5 Deep Think, a super-high-performance math system

Google has launched the Gemini 2.5 Deep Think, a machine with enhanced reasoning capabilities, for subscribers of its top-tier AI subscription service, Google AI Ultra. Its unique feature is its ability to solve mathematical problems.

Gemini 2.5: Deep Think is now rolling out

'Gemini 2.5 Deep Think' is a version of the model that achieved gold medal standards at the 2025 International Mathematical Olympiad.

Google's enhanced Gemini reaches the gold medal performance at the Mathematical Olympiad, works in natural language, and derives answers within the same time limit as humans - GIGAZINE

While the above models are strong in mathematics, they take several hours to infer complex problems. The newly released Gemini 2.5 Deep Think has reduced performance to the level of a bronze medal winner at the International Mathematical Olympiad, but has shortened inference time to optimize it for everyday use. A gold medal-level model has also been officially released.

Gemini 2.5 Deep Think utilizes a technique called 'parallel thinking,' which allows it to generate many ideas simultaneously and consider them in parallel. It then takes time to refine and combine ideas before arriving at the best answer. This allows it to deliver detailed, creative, and thoughtful responses.

The newly released model excels at iterative development, tasks that involve incrementally building complex structures, formulating and exploring mathematical hypotheses, analyzing complex scientific literature, and solving coding problems. In benchmarks that test reasoning, knowledge, coding, and mathematical abilities, it outperforms Gemini 2.5 Pro, OpenAI o3, and Grok 4, which also have reasoning capabilities.

Gemini 2.5 Deep Think also demonstrated improved safety and tone objectivity compared to Gemini 2.5 Pro, but was slightly more likely to reject harmless requests.

Google AI Ultra subscribers can take advantage of Gemini 2.5 Deep Think by selecting 2.5 Pro as their model and turning on 'Deep Think' in the prompt bar. Deep Think automatically integrates with tools like code execution and Google search to generate longer responses.

In the coming weeks, we will release Deep Think with and without the tooling to trusted testers via the Gemini API, and continue our work to better understand usability for developers and enterprises.