Google Rolls Out Gemini 3 Deep Think to AI Ultra Users

Google has officially begun rolling out its most advanced reasoning capability, Gemini 3 Deep Think, to subscribers of its top-tier "Google AI Ultra" plan. The move introduces a powerful new "parallel reasoning" system designed to tackle complex mathematical, scientific, and coding challenges, positioning it as a direct competitor to OpenAI’s o1 model.

The rollout, which began earlier this week, marks a significant delineation in Google’s AI product strategy, reserving its most computationally expensive "thinking" features for its highest-paying customers.

What is "Deep Think"?
Unlike standard chatbots that predict the next word in a sentence, Deep Think is engineered to pause and deliberate before responding. It utilizes a technique Google calls "advanced parallel reasoning," which allows the model to explore multiple hypotheses simultaneously rather than following a single linear path.

"Deep Think uses iterative rounds of reasoning to explore multiple paths at once," Google explained in its release notes. "This results in more refined, nuanced code and the ability to solve problems that stump standard models."

According to Google, the model has achieved record-breaking scores on industry benchmarks:

41.0% on "Humanity’s Last Exam" (a test designed to be too difficult for current AI).

93.8% on GPQA Diamond (PhD-level science questions).

45.1% on ARC-AGI-2 (a visual reasoning test) when equipped with code execution tools.

Exclusive to the $250 "Ultra" Tier
Access to the full "Deep Think" capability comes at a steep price. The feature is currently exclusive to the Google AI Ultra plan, a new subscription tier priced at $249.99 per month.

While Google’s standard $19.99/month "AI Pro" plan includes a feature called "Thinking with Gemini 3 Pro" (limited to roughly 100 prompts a day), the dedicated "Deep Think" mode—which offers deeper search depth and a specialized 192,000-token context window for heavy reasoning—is reserved for Ultra subscribers.

In addition to Deep Think, the Ultra plan includes massive perks like 30TB of cloud storage, early access to the Veo 3 video generation model, and the new Project Mariner autonomous agent prototype.

A New "Thinking" Interface
For eligible users, the feature appears as a new "Deep Think" toggle in the prompt bar of the Gemini app.

When activated, the interface changes to show the model's "thought process" in real-time, often taking several minutes to generate a final answer. This transparency is designed to help researchers and developers understand how the AI arrived at its conclusion—a critical feature for high-stakes fields like engineering and data science.

This release escalates the "reasoning race" between tech giants, as companies shift focus from simple chatbots to "System 2" thinkers capable of long-horizon planning and problem-solving.