Key takeaways Google has begun rolling out Gemini 3 Deep Think mode to users on the Google AI Ultra subscription. The new mode significantly enhances reasoning capabilities, enabling complex math, sci...
On December 4, 2025, Google officially rolled out Gemini 3 Deep Think to subscribers of Google AI Ultra.
The update, released through the Gemini app, brings the company’s most advanced reasoning mode yet to paying users.
Through Deep Think, the system applies what Google describes as “parallel reasoning,” evaluating multiple hypotheses simultaneously to tackle complex mathematical, scientific, and logic-based problems.
In a statement on its release notes, Google explained that Deep Think uses “iterative rounds of reasoning to explore multiple hypotheses simultaneously” to solve problems and is designed to help users “tackle even more complex problems.”
Users with Ultra subscriptions can access the feature by opening the Gemini app, selecting “Deep Think” in the prompt bar, and choosing “Thinking with 3 Pro” from the model dropdown.
According to Google, Gemini 3 Deep Think outperforms not only earlier versions of Gemini but also many leading AI models across benchmark tests designed to measure deep reasoning, scientific knowledge, visual reasoning, and tool use.
On the Humanity’s Last Exam benchmark, measuring reasoning and knowledge, Deep Think achieves a score of 41.0%, outperforming Gemini 3 Pro’s 37.5%.
On GPQA Diamond (scientific knowledge), it records 93.8%.
On ARC-AGI-2 (visual reasoning with code execution), it scores 45.1%, marking a substantial improvement over prior capabilities.
These gains build on the advances of Gemini 3, released November 18, 2025, which already expanded multimodal understanding (text, image, video, audio, code), afforded a 1-million-token context window, and enabled improved tool use, long-form reasoning, and coding tasks.
Together, Gemini 3 enhancements, now including Deep Think, are described by Google as a “step-change” in what their AI can handle.
The limited rollout only to Ultra subscribers suggests a strategy to reserve the most advanced reasoning features for users willing to pay a premium.
By focusing on high-value customers first, Google may be using the rollout as a controlled test of Deep Think’s safety, robustness, and real-world performance before wider release.
In a Google blog post accompanying the release, the company said: “We’re committed to bringing our latest AI innovations faster to Google AI Ultra users, including our most powerful reasoning capabilities to help you tackle even more complex problems.”
For now, Deep Think remains unavailable to free-tier users, reflecting a trend across the industry of gating top-tier AI capabilities behind subscription plans.
For users with access, Gemini 3 Deep Think represents a powerful new tool for tasks that demand deep reasoning: from advanced mathematics, logic puzzles, scientific analysis, to multi-step coding challenges.
Its ability to reason across multiple modalities, including code execution and visual reasoning, may expand the kinds of problems AI can help solve, perhaps bringing AI closer to assisting with research, planning, and technical work.
For developers and researchers, the release is notable because it signals that AI is moving beyond pattern matching and surface-level text generation, toward more abstract reasoning, hypothesis testing, and tool-driven workflows.
That said, the restriction to a premium tier means broad access remains limited, raising questions about equity and the long-term role of powerful AI tools in society.
As AI models become more capable, Google’s strategy suggests it will continue to roll out advanced features first to paying users, likely refining and testing before any broader public availability.
Mistral AI Releases Family Of Open Models Targeting Enterprise And Edge Computing
Microsoft Faces Scrutiny Over AI Product Sales Amid Customer Resistance
Thai Police Apologize For AI-Manipulated Flood Relief Image
Microsoft is reportedly exploring OpenClaw-style AI features for Copilot that could make the assistant more proactive inside Microsoft 365.
Meta launches Muse Spark, a closed proprietary AI model with tiered reasoning and 3B+ user reach across WhatsApp, Instagram, and Quest VR.
Key takeaways OpenAI has signed a multi-year, $10 billion agreement with AI chipmaker Cerebras Systems to secure computing infrastructure. The deal will deliver 750 megawatts of computing power throug...
