Google DeepMind today began rolling out Gemini 2.5 Deep Think, a reasoning model that uses parallel thinking to generate and evaluate multiple ideas at once before converging on an answer. The model is available in the Gemini app for Google AI Ultra subscribers ($250/month), switched on with an optional “Deep Think” toggle when using 2.5 Pro.

Deep Think is a variation of the system that recently scored a gold medal at the 2025 International Mathematical Olympiad. That IMO model takes hours to reason; the shipped version reaches roughly bronze-level IMO performance while running faster. Google says it leads on benchmarks including Humanity’s Last Exam (34.8% without tools) and LiveCodeBench V6 (87.6%), ahead of OpenAI’s o3 and xAI’s Grok 4. The company also developed novel reinforcement learning techniques to encourage the model to make better use of extended reasoning paths.

Google is simultaneously giving a small group of mathematicians and academics access to the full IMO-grade model, hoping to gather feedback for further development. Google says that in testing, Gemini 2.5 Deep Think showed improved content safety and tone-objectivity compared with Gemini 2.5 Pro, but also had a higher tendency to refuse benign requests. Over the coming weeks, Google will also make the model available to select testers via the Gemini API.

The move comes as several leading AI labs are converging on multi-agent systems, while Google and xAI are choosing to gate some of them behind expensive subscription plans.