Google has introduced Deep Think mode for Gemini 2.5 Pro — an advanced computing mode that allows the model to consider multiple answer options simultaneously before choosing a final version.
Google has introduced Deep Think mode for Gemini 2.5 Pro — an advanced computing mode that allows the model to consider multiple answer options simultaneously before choosing a final version.
According to The Verge, this approach uses multi-agent technology: the model generates several ideas in parallel, analyzes and combines them, and only then forms a result. This technique significantly improves the quality of problem solving — especially in mathematics, programming, and strategic thinking.
At the International Mathematical Olympiad (IMO) 2025, Gemini Deep Think solved 5 out of 6 problems and received a gold medal — 35 out of 42 points, as assessed by the official IMO coordinators.
In the LiveCodeBench V6 test (for programming competitions), Gemini 2.5 Deep Think outperformed other AI competitors, including OpenAI o3 and Grok 4.
On the complex Humanity’s Last Exam (HLE) benchmark, the model achieved 34,8% accuracy, outperforming xAI Grok 4 (~25%) and OpenAI o3 (~20%).
Deep Think mode is available to Google AI Ultra subscribers (≈ $250/month) via the Gemini mobile app and API.
Google is distributing a version of the model that showed results on IMO to a limited group of scientists to get feedback before scaling up.
Deep Think uses methods of extended inference or «thinking time» — the model spends more time on internal processing, «thinks» about several options, compares, combines, and only then gives an answer.
The company is also applying new reinforcement learning techniques to make the model better utilize multi-step thinking and evolve over time.
Google positions Deep Think as a priority model for tasks that require a creative approach, strategic planning, and gradual improvement, such as design creation, algorithmic challenges, scientific research, or the development of code agents.
Demis Hassabis, head of DeepMind, noted: many of the acquired thinking capabilities underlie the potential development of AGI (artificial general intelligence) — especially the ability of a model to reconfigure, contextualize, and act creatively in the real world.
