Google Launches Gemini 2.5 Deep Think for AI Ultra Users
▼ Summary
– Google is launching its most advanced Gemini model, Gemini 2.5 Deep Think, but access is limited to subscribers of the $250 AI Ultra plan.
– Deep Think is optimized for complex queries, requiring more compute resources and offering enhanced parallel analysis for deeper problem-solving.
– The model explores multiple approaches to problems, revisiting and remixing hypotheses to generate higher-quality outputs.
– Deep Think excels in design aesthetics, scientific reasoning, and coding, though it takes several minutes to produce answers.
– Benchmark tests show Deep Think outperforms Gemini 2.5 Pro and rivals like OpenAI o3 and Grok 4, particularly in Humanity’s Last Exam with a 34.8% score.
Google has rolled out its most advanced AI model yet, Gemini 2.5 Deep Think, though access remains exclusive to high-tier subscribers. The cutting-edge system, first unveiled at Google I/O earlier this year, is now live in the Gemini app but limited to users on the premium $250 AI Ultra plan. Designed for tackling intricate problems, this iteration demands significantly more computational power than standard versions, justifying its restricted availability.
Built on the same framework as Gemini 2.5 Pro, Deep Think distinguishes itself by extending “thinking time” and employing parallel analysis to explore multiple solution pathways. Google explains that the model doesn’t just generate answers, it revisits, refines, and even combines hypotheses to produce higher-quality results. This methodical approach enhances its performance in specialized areas like design aesthetics, scientific reasoning, and complex coding tasks.
Unlike faster, lightweight models, Deep Think isn’t built for speed. Responses can take several minutes, reflecting the system’s deeper analytical process. In benchmark testing, it outperformed not only its Pro counterpart but also rival models such as OpenAI’s o3 and Grok 4. One standout achievement was its performance on Humanity’s Last Exam, a rigorous assessment featuring 2,500 multi-disciplinary questions spanning over 100 subjects. While competitors scored between 20-25%, Deep Think achieved an impressive 34.8% accuracy, showcasing its superior problem-solving capabilities.
For now, this powerhouse remains a niche tool, reserved for those willing to invest in top-tier AI performance. Whether broader access follows depends on Google’s strategy for balancing resource demands with user demand.
(Source: Ars Technica)
