Deep Think, a multi-agent AI model from the Google DeepMind team, is now available via the Gemini app. While this is a more accessible variation of the model that recently achieved the gold standard at the International Math Olympiad (IMO) 2025, the full-strength version is also being made available to select academics and researchers.
Dual rollout for different user groups
Starting on Aug. 1, users who have an active Google AI Ultra subscription can access Deep Think at no additional cost. However, this isn’t the same version that was used in the IMO 2025. That high-performing version can reason over several hours when solving complex math problems, which makes it less practical for everyday or commercial use.
Instead, Google AI Ultra subscribers will receive a version of Deep Think designed for faster performance. Although it does not replicate the gold-standard performance of the research-grade version, it still demonstrates strong reasoning capabilities and ranks higher than competing models in standard evaluations.
The research variant is not being shelved. Google is releasing that version — albeit to a limited group of academic researchers and mathematicians — to support advanced testing and further optimization.
Answering questions from multiple perspectives
The strength of Deep Think lies in its multi-agent architecture. By deploying several reasoning agents in parallel, Deep Think can process multiple ideas and concepts simultaneously. It then selects, refines, or merges the most promising threads into a final response.
A recent blog post by the Deep Think team reads, in part: “Just as people tackle complex problems by taking the time to explore different angles, weigh potential solutions, and refine a final answer, Deep Think pushes the frontier of thinking capabilities by using parallel thinking techniques. This approach lets Gemini generate many ideas at once and consider them simultaneously, even revising or combining different ideas over time, before arriving at the best answer.”
Examining the benchmarks
Despite not being the IMO-grade version, the Deep Think model offered to Ultra subscribers still performs strongly in benchmark tests, outpacing other leading AI models like Gemini 2.5 Pro.
| Reasoning & Knowledge (Humanity’s Last Exam) | Code Generation (LiveCodeBench) | Mathematics (IMO 2025) | Mathematics (AIME 2025) | |
|---|---|---|---|---|
| Gemini 2.5 Deep Think | 34.8% | 87.6% | 60.7% | 99.2% |
| Gemini 2.5 Pro | 21.6% | 74.2% | 31.6% | 88.0% |
| OpenAI-o3 | 20.3% | 72.0% | 16.7% | 88.9% |
| Grok-4 | 25.4% | 79.0% | 21.4% | 91.7% |
Optimizing Deep Think even further
Two more versions of Deep Think — one with tools and one without — will soon be released to trusted testers and researchers via the Gemini API. With several model versions in circulation, Google DeepMind aims to collect broader usage insights and refine its future AI models to support both commercial and research-level applications.
Want to understand what Microsoft’s new AI hires mean for the future of deep learning? Don’t miss our report on how the tech giant is quietly reshaping its AI research stack.
