Google has released its advanced Deep Think model in the Gemini app, making it available to all Google AI Ultra subscribers as of August 1st. The rollout follows internal performance gains and user feedback from trusted testers.

The version now available is a faster, more accessible iteration of the model that reached gold-medal-level performance at the 2025 International Mathematical Olympiad (IMO). While the competition model required hours of reasoning, this public release maintains Bronze-level IMO performance with significantly reduced response time.

To support early research applications, Google is also providing access to the full Gemini 2.5 Deep Think model—the version submitted to the IMO—to a small group of mathematicians and academics. One of the early testers is mathematician Michel van Garrel, who explored mathematical conjectures using the model.

“We look forward to hearing how it could enhance their research and inquiry, and we’ll use their feedback as we continue to improve this offering,” Google stated in a blog announcement.

Deep Think extends the Gemini architecture by using parallel thinking techniques, enabling it to generate and weigh multiple ideas at once. This process mimics how humans approach complex problems—by evaluating and refining ideas over time.

The model benefits from longer inference times, or “thinking time,” allowing it to explore a broader range of hypotheses before arriving at conclusions. Google has paired this with new reinforcement learning techniques to help the model improve its multi-step reasoning over time.

According to internal tests, Deep Think achieves state-of-the-art results in several advanced benchmarks, including:

The model is designed to assist with tasks requiring creativity and strategic planning, such as:

As with other Gemini models, Deep Think is being released with a focus on safety. Google reports that Gemini 2.5 Deep Think shows improved content safety and tone objectivity compared to Gemini 2.5 Pro, although it may be more likely to refuse benign requests.

The company continues to apply frontier safety evaluations and has planned mitigations in place for higher capability levels. More details are available in the model card for Gemini 2.5 Deep Think.

AI Ultra subscribers can access Deep Think today in the Gemini app by selecting 2.5 Pro from the model dropdown and toggling the “Deep Think” option in the prompt bar. The tool supports longer responses and works seamlessly with Google Search and code execution tools.

In addition, Google plans to release versions of Deep Think—both with and without tools—via the Gemini API to trusted testers in the coming weeks, as part of its effort to explore enterprise and developer applications.

Q: What is Deep Think?
A: Deep Think is a new AI reasoning model from Google, available through the Gemini app, designed to solve complex problems using parallel thinking and extended inference time.

Q: Who can access Deep Think?
A: It’s available to all Google AI Ultra subscribers, with additional early access to a full version for select mathematicians and researchers.

Q: What is Deep Think used for?
A: It supports iterative design, scientific discovery, and advanced coding, making it valuable for creative and technical problem solving.

Q: How does Deep Think perform in benchmarks?
A: The model achieves state-of-the-art results in LiveCodeBench V6 and Humanity’s Last Exam, outperforming models without tool use.

Q: How can users activate Deep Think in the Gemini app?
A: By selecting 2.5 Pro in the model dropdown and toggling “Deep Think” in the prompt bar.

The release of Deep Think marks a significant advancement in Gemini’s reasoning capabilities, especially in areas that demand creative, multi-step problem solving. By combining parallel thinking, longer inference times, and reinforcement learning, Deep Think brings Google’s most advanced approach to reasoning directly into the hands of AI Ultra subscribers—a move that aligns with a growing industry focus on deeper, more reflective AI behavior.

While other models from OpenAI and Anthropic also aim to improve reasoning with features like tool use or multi-agent collaboration, Deep Think stands out for its emphasis on sustained, parallel reasoning paths and its early performance on math and code-intensive benchmarks. These strengths make it particularly valuable for tasks involving scientific discovery, algorithm development, and step-by-step design thinking.

That said, the current form of Deep Think is clearly optimized for researchers, developers, and technical users. It’s not yet designed for casual, everyday use—interactions are limited to a fixed number of prompts, and activating it requires toggling settings within the Gemini app’s model menu. Broader adoption may depend on how well it can generalize to less technical domains or assist with open-ended, non-coding tasks.

Still, Deep Think’s debut signals a shift toward more deliberative AI systems that can extend human reasoning rather than simply respond quickly. For teams working on hard problems—whether in academia, enterprise, or engineering—it could mark the beginning of a new kind of collaboration, where AI doesn't just assist but thinks alongside us.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.