Google officially launches Gemini 2.5 Deep Think: Olympic-level mathematical AI now in the hands of users
On August 1, 2025, Google officially announced the Deep Think feature in its Gemini 2.5 model, now available to Gemini Advanced subscribers. This marks a new milestone, as a commercial AI model demonstrates high-level mathematical problem-solving capabilities, reaching bronze level in the IMO-standard simulation test developed by Google.
What is Deep Think?
Deep Think is an advanced reasoning mode integrated into Gemini 2.5 - Google’s latest and most powerful language model. Unlike traditional AI models that respond based on a single linear chain of thought, Deep Think operates using a “parallel thinking” approach.
Specifically, when receiving a question, Deep Think will:
- Simultaneously generate multiple reasoning paths, hypotheses and potential solutions;
- Compare, evaluate and select the most appropriate solution before responding to the user.
This approach mimics the thinking process of human experts, enabling the model to tackle complex logic problems, multi-step mathematical tasks and advanced coding - instead of merely generating surface-level text.
It can be said that Gemini 2.5 Deep Think is the first AI system to approach Olympic-level problem solving.
One of the clearest demonstrations of Deep Think’s capabilities is its success in solving International Mathematical Olympiad (IMO) problems. In internal testing, the research version of Gemini 2.5 Deep Think scored 35/42 points - equivalent to a gold medal according to IMO standards.
Although this gold-level version is currently used internally at Google DeepMind, users of the Ultra subscription tier can now experience an optimized version suitable for real-world use, achieving bronze-level performance on equivalent benchmarks.
What’s included in the public release?
Users of the Google AI Ultra plan ($249/month) can now access Deep Think via:
- The Gemini app (mobile or web);
- Select the Gemini 2.5 Pro model → enable the Deep Think option in Settings.
The released version still retains:
- Multi-step reasoning capabilities, especially in mathematics, physics and coding;
- Integration with Google Search, code execution and long, well-structured answers;
- A simple, intuitive interface (responses may be slower, but are deeper and better explained).
In addition to IMO, Deep Think also performs strongly on other high-standard benchmarks:
- LiveCodeBench v6 – focused on advanced code generation and debugging;
- Humanity’s Last Exam – a set of questions designed to test reasoning, ethics and critical thinking;
- Outperforming models such as GPT-4o, Claude 3.5 Sonnet and Grok 4 in reasoning depth and structured output metrics.
As Sundar Pichai noted on the Google Blog, Deep Think is an important step toward bringing AI closer to human-level understanding, reasoning and collaboration. Google also emphasized that Deep Think will become a key component on the path toward Artificial General Intelligence (AGI), with even more powerful versions currently being tested in areas such as education, mathematics and open science.
In short, Deep Think is not just a feature update - it is a statement that AI can truly become a thinking partner, accompanying humans in solving the most challenging problems.
Sources:
- Google. (2025, August 1). Gemini 2.5: Deep Think is now rolling out. Google Blog. https://blog.google/products/gemini/gemini-2-5-deep-think/
- TechCrunch. (2025, August 1). Google rolls out Gemini Deep Think AI, a reasoning model that tests multiple ideas in parallel. https://techcrunch.com/2025/08/01/google-rolls-out-gemini-deep-think-ai-a-reasoning-model-that-tests-multiple-ideas-in-parallel/
- Tom’s Guide. (2025, August 1). Google’s new Deep Think feature is here: What it does and why it might not stay Ultra-exclusive for long. https://www.tomsguide.com/ai/googles-new-deep-think-feature-is-here-what-it-does-and-why-it-might-not-stay-ultra-exclusive-for-long
