Google Gemini Deep Think AI: Parallel Reasoning Model Released

Google Unveils Gemini 2.5 Deep Think: An Advanced AI Reasoning Model

Google DeepMind has initiated the rollout of Gemini 2.5 Deep Think, representing its most sophisticated AI model to date. This new system is designed to tackle complex questions by simultaneously exploring a multitude of ideas.

The model then leverages these diverse outputs to determine the optimal response, showcasing a significant leap in AI reasoning capabilities.

Availability and Subscription Details

Access to Gemini 2.5 Deep Think will be granted to subscribers of Google’s Ultra plan, priced at $250 per month, beginning this Friday within the Gemini app.

The Multi-Agent Approach

Initially presented in May at Google I/O 2025, Gemini 2.5 Deep Think marks Google’s inaugural publicly accessible multi-agent model.

These systems function by generating multiple AI agents to address a single query in parallel. While this process demands substantial computational resources, it generally yields superior results.

Success in Competitive Environments

Google successfully employed a variant of Gemini 2.5 Deep Think to achieve a gold medal at the recent International Math Olympiad (IMO).

Furthermore, the specific model utilized during the IMO is being released to a limited cohort of mathematicians and academics.

Research and Development Focus

Google emphasizes that this IMO model requires “hours to reason,” contrasting with the seconds or minutes typical of consumer AI models. The company intends for this release to foster research and gather feedback for refining the multi-agent system for academic applications.

Improvements Over Initial Release

Google asserts that the current Gemini 2.5 Deep Think model represents a considerable advancement over the version showcased at I/O.

They have also reportedly developed “novel reinforcement learning techniques” to optimize the model’s utilization of its reasoning pathways.

Potential Applications

“Deep Think can help people tackle problems that require creativity, strategic planning and making improvements step-by-step,” Google stated in a blog post shared with TechCrunch.

Performance Benchmarks

Gemini 2.5 Deep Think has demonstrated state-of-the-art performance on Humanity’s Last Exam (HLE), a rigorous test of AI’s ability to answer a broad range of questions.

The model achieved a score of 34.8% on HLE (without tools), surpassing xAI’s Grok 4 (25.4%) and OpenAI’s o3 (20.3%).

Additionally, Gemini 2.5 Deep Think outperformed models from OpenAI, xAI, and Anthropic on LiveCodeBench 6, a challenging coding competition, scoring 87.6% compared to 79% for Grok 4 and 72% for o3.

Enhanced Capabilities

Gemini 2.5 Deep Think seamlessly integrates with tools like code execution and Google Search, and is capable of generating “much longer responses” than conventional AI models.

Testing by Google revealed that the model produces more detailed and visually appealing results for web development tasks compared to its competitors, potentially accelerating research and discovery.

Industry Trend Towards Multi-Agent Systems

A convergence towards the multi-agent approach is becoming apparent among leading AI laboratories.

xAI’s Grok 4 Heavy, recently released by Elon Musk’s company, also utilizes a multi-agent system and has demonstrated leading performance on several benchmarks.

Similarly, OpenAI’s model used to secure a gold medal at the IMO was also a multi-agent system, as confirmed by researcher Noam Brown. Anthropic’s Research agent, known for its comprehensive research briefs, is also powered by a multi-agent system.

Cost Considerations

Despite their strong performance, multi-agent systems are demonstrably more expensive to operate than traditional AI models.

Consequently, technology companies may restrict access to these systems through their premium subscription tiers, a strategy adopted by both xAI and now Google.

Future Plans

In the coming weeks, Google intends to share Gemini 2.5 Deep Think with a select group of testers via the Gemini API.

This initiative aims to gain a deeper understanding of how developers and enterprises might leverage the capabilities of its multi-agent system.