Google Launches Gemini 3 Deep Think for Ultra Subscribers, Showcases Major Leaps in Reasoning and AGI Benchmarks

Highlights

• Google has introduced Gemini 3 Deep Think, a reasoning-focused AI model, available exclusively to AI Ultra subscribers at Rs. 24,500/month.
• The model uses a new parallel reasoning technique, exploring multiple solution paths simultaneously and refining results for more coherent answers.
• Gemini 3 Deep Think scored 41% on Humanity’s Last Exam and 45.1% on ARC-AGI-2.

Google has officially rolled out its Gemini 3 Deep Think model to select users, expanding its advanced AI lineup just weeks after the global launch of Gemini 3 Pro. It is positioned as Google’s next-generation reasoning-focused model and aims to tackle highly complex and niche problems across logic, mathematics, science, and multi-step reasoning. However, access to the model is restricted to the company’s most premium subscription tier.

Gemini 3 Deep Think Now Available for AI Ultra Subscribers

In a blog post, Google confirmed that the new large language model (LLM) has been released exclusively for users subscribed to the Google AI Ultra plan. The company described the model as designed for problems that “challenge even the most advanced state-of-the-art models.”

Only a small group of users paying Rs. 24,500 per month for the Google AI Ultra subscription can access the model. They can enable it through the Gemini website or app by choosing “Deep Think” in the prompt bar and selecting Gemini 3 Pro from the model picker.

Powered by New ‘Parallel Reasoning’ Technique

Google says Gemini 3 Deep Think’s performance is driven by a new method called parallel reasoning. Unlike traditional models that follow a single chain of thought, Deep Think evaluates multiple solution paths simultaneously. It then refines the results through iterative rounds and selects the most coherent final answer.

Interest in Gemini 3 Deep Think has surged since its initial mention last month, largely driven by its standout benchmark results. Google highlighted two tests in particular that contributed to the model’s early attention.

The first is Humanity’s Last Exam, a notoriously tough academic reasoning benchmark. Google claims Gemini 3 Deep Think scored 41 per cent without tools, the highest score to date and significantly above other AI models.

The second is the ARC-AGI-2 benchmark, an advanced reasoning test focused on visual puzzles that require code execution for solutions. Google reports a score of 45.1 per cent, calling the achievement “unprecedented.” Notably, the test is also viewed as an indicator of a model’s progress toward artificial general intelligence (AGI).

FAQs

Q1. Who can access Google’s Gemini 3 Deep Think model?

Answer. Gemini 3 Deep Think is available exclusively to Google AI Ultra subscribers, costing Rs. 24,500 per month.

Q2. What makes Gemini 3 Deep Think different from other AI models?

Answer. It uses a new parallel reasoning technique, evaluating multiple solution paths simultaneously and refining results to deliver more coherent answers.

Q3. How has Gemini 3 Deep Think performed in benchmarks?

Answer. The model scored 41% on Humanity’s Last Exam and 45.1% on ARC-AGI-2, both unprecedented results that highlight progress toward advanced reasoning and AGI.

Also Read –

https://www.mymobileindia.com/google-adds-new-tool-to-gemini-app-allowing-users-to-check-whether-an-image-was-created-with-ai/

https://www.mymobileindia.com/google-launches-nano-banana-pro-with-gemini-3-pro-enhanced-ai-image-generation-now-supports-4k-output/