Google Launches Gemini Deep Think AI for Parallel Reasoning and Idea Testing

Google DeepMind introduces Gemini 2.5 Deep Think, their most sophisticated AI reasoning model to date. This innovative system can analyze multiple thought pathways concurrently before selecting the optimal solution.

Starting Friday, subscribers to Google's premium $250/month Ultra plan will receive exclusive access to Gemini 2.5 Deep Think through the Gemini app.

Debut at Google I/O 2025, Gemini 2.5 Deep Think represents Google's inaugural public multi-agent model. The system deploys numerous AI agents simultaneously to address complex queries, requiring greater computational power but yielding superior results.

A specialized version of Gemini 2.5 Deep Think secured a gold medal at this year's International Math Olympiad.

Google is sharing its IMO-winning model with mathematicians and academics while developing Gemini 2.5 Deep Think. Unlike conventional AI models that respond instantly, this research-focused version requires hours of reasoning time. The company aims to refine multi-agent technology for academic applications through this initiative.

Google highlights significant advancements in Gemini 2.5 Deep Think since its initial I/O announcement, including novel reinforcement learning techniques that optimize the model's reasoning capabilities.

"Deep Think excels at solving problems demanding creative thinking, strategic planning, and iterative refinement," Google stated in a TechCrunch-exclusive blog post.

Tech and VC heavyweights join the Disrupt 2025 agenda

Industry leaders including Netflix, ElevenLabs, Wayve, and Sequoia Capital headline TechCrunch Disrupt 2025. Gain transformative insights to propel your startup's growth at our milestone 20th anniversary event. Secure your tickets today for exclusive savings up to $675 before prices increase.

Tech and VC heavyweights join the Disrupt 2025 agenda

Industry leaders including Netflix, ElevenLabs, Wayve, and Sequoia Capital headline TechCrunch Disrupt 2025. Gain transformative insights to propel your startup's growth at our milestone 20th anniversary event. Secure your tickets today for exclusive savings up to $675 before prices increase.

Gemini 2.5 Deep Think sets new standards on Humanity's Last Exam (HLE), a comprehensive evaluation testing AI across multiple disciplines. Google reports 34.8% accuracy without external tools, surpassing competitors Grok 4 (25.4%) and OpenAI's o3 (20.3%).

The model also dominates LiveCodeBench 6 with an 87.6% score, outperforming Grok 4 (79%) and OpenAI's o3 (72%) in competitive programming tasks.

Gemini 2.5 Deep Think seamlessly integrates with programming environments and search tools while generating more extensive outputs than conventional AI systems.

Google's internal tests show the model produces exceptionally detailed web development solutions, potentially accelerating scientific breakthroughs.

Leading AI researchers increasingly adopt multi-agent architectures.

xAI's Grok 4 Heavy and OpenAI's unreleased IMO model both utilize multi-agent systems, as does Anthropic's research-oriented AI. While these systems demonstrate superior performance, their operational costs necessitate premium pricing tiers.

Google plans limited API access for Gemini 2.5 Deep Think in upcoming weeks to explore enterprise and developer applications.