Google launches Gemini 3 Deep Think, its most advanced reasoning AI for complex problem-solving

Google rolling out Gemini 3 Deep Think to AI Ultra

Deep Think is Google's "most advanced reasoning mode" that leverages an advanced parallel reasoning technique to "explore multiple hypotheses simultaneously." These "iterative rounds of reasoning" can result in "more refined, nuanced code" for better visualizations and prototyping. This is for complex math, science, and logic problems, with answers usually taking a few minutes to come back. Building on the Gemini 2.5 version introduced in August, Deep Think has marked improvements in the following benchmarks: In mid-November, Google said it needed "extra time for safety evaluations and input from safety testers." Google AI Ultra subscribers ($250/month) can now access "Deep Think" from the Tools menu when "Thinking" is set in the model dropdown.

Google just launched Gemini 3 Deep Think -- its most powerful AI model yet

Google still isn't done with its Gemini 3 rollout. After weeks of new features, it still has some cards up its sleeve, now announcing the launch of 'Deep Think'. This is a new thinking mode, coming to the Gemini app, exclusively for those on the Google AI Ultra plan. As the name suggests, it's a version designed for pushing Gemini harder on more complex tasks. "This new mode delivers a meaningful improvement in reasoning capabilities, designed to tackle complex math, science and logic problems that challenge even the most advanced state-of-the-art models," wrote Tulsee Doshi, Senior Director, Product Management at Gemini, in a blog post. Exact details of what this model provides are fairly slim right now. Google claims that it is industry leading on complicated benchmarks, such as Humanity's Last Exam and ARC-AGI 2. These are tests that examine a chatbot's ability in logic and reasoning, seeing how well it can deal with challenges that can normally cause problems for chatbots. Google claims that its ability to do so well on these types of tests is due to it using advanced parallel reasoning. That, essentially, means that it can explore multiple hypotheses simultaneously to generate the correct answer faster. As mentioned above, Gemini 3 Deep Think is only available for AI Ultra subscribers. That is Google's most expensive AI plan, but it unlocks access to everything Gemini has to offer. If you are on the AI Ultra plan, simply select 'Deep Think' from the dropdown bar in the Gemini chatbox. With this selected, you can now prompt using the smarter model of Gemini 3. This is a term doing the rounds a lot right now in AI. Thinking, or reasoning models, as they are also referred to, are simply versions of AI tools designed to take on more challenging tasks. They are trained to mimic the way humans solve problems, breaking through complex tasks via logical processes of thought. By doing this, the model can give more time to a task, either generating a more thought-out response, or being able to work out a problem it wouldn't normally understand. AI models often come in different levels of this. Tools like ChatGPT, Gemini and Claude offer levels starting off with versions of their tools that are cheaper to run and faster, but aren't able to deal with complicated tasks. While for a lot of people, models like Gemini 3's Deep Think will offer more power than needed, these are the versions you hear about in the news, breaking benchmarking tests and outperforming humans.

Gemini 3 Deep Think rolls out to paid subscribers

After launching Gemini 3 in late November, Google is rolling out its latest Deep Think mode to AI Ultra subscribers. Now available in the Gemini app for anyone shelling out $250 a month for the pro plan, Gemini 3 Deep Think is the supercharged mode of Google's latest large language model (LLM), designed to tackle more complex problems. According to the company, it "uses advanced parallel reasoning to explore multiple hypotheses simultaneously." The latest update after Gemini 2.5, Gemini 3 Deep Think has a few bragging rights, as Google says it scored 41 percent on the extremely difficult LLM benchmark, Humanity's Last Exam. If you're paying for AI Ultra, you can try Gemini 3 Deep Think in the Gemini app by hitting the prompt bar and selecting the dropdown option. Demand is high for Google's AI tools of late, despite the company joining competitors like Anthropic and OpenAI in taking home an average AI safety report card this week. The rollout comes days after Google had to cap access to its next gen image generator, Nano Banana, temporarily dropping the number of free image prompts from three to two (this didn't affect Google AI Pro or AI Ultra plan subscribers). Google says its Gemini app is used by 650 million users every month, and with the arrival of Gemini 3, rival OpenAI reportedly lost 6 percent of its users in a week.

Gemini 3 Deep Think is now available in the Gemini app.

Today, we're rolling out Gemini 3 Deep Think mode to Google AI Ultra subscribers in the Gemini app. This new mode delivers a meaningful improvement in reasoning capabilities, designed to tackle complex math, science and logic problems that challenge even the most advanced state-of-the-art models. Gemini 3 Deep Think excels on rigorous benchmarks like Humanity's Last Exam and ARC-AGI-2. This is because it uses advanced parallel reasoning to explore multiple hypotheses simultaneously -- building on Gemini 2.5 Deep Think variants that recently achieved a gold-medal standard at the International Mathematical Olympiad and at the International Collegiate Programming Contest World Finals. Ultra subscribers can try Gemini 3 Deep Think mode today by selecting "Deep Think" in the prompt bar and Gemini 3 Pro in the model dropdown.

Google Launches Gemini 3 Deep Think Mode for Ultra Subscribers | AIM

The system uses parallel reasoning to test multiple hypotheses, building on earlier Deep Think variants. Google has introduced Gemini 3 Deep Think mode within the Gemini app, offering advanced reasoning capabilities for complex problem-solving. "Gemini 3 Deep Think is now available in the Gemini app," wrote Tulsee Doshi, senior director of product management, in an announcement on December 4. The new mode is available to Google AI Ultra subscribers and is designed to handle complex math, science, and logic queries. Users can access the feature by selecting "Deep Think" in the prompt bar and choosing Gemini 3 Pro from the model menu. According to Google, Gemini 3 Deep Think demonstrates strong performance on benchmarks such as Humanity's Last Exam, where it scored 41% without tools, and ARC-AGI-2, where it reached 45.1% with code execution. The system uses parallel reasoning to test multiple hypotheses, building on earlier Deep Think variants that performed well in competitions like the International Mathematical Olympiad and the ICPC World Finals. Google recently launched the broader Gemini 3 lineup, positioning it as a step toward more capable AI systems. The company said Gemini 3 Pro outperforms Gemini 2.5 Pro, OpenAI's GPT-5.1, and Anthropic's Claude Sonnet 4.5 across benchmarks, including LMArena, Humanity's Last Exam, GPQA Diamond, and MathArena Apex. With expanded multimodal inputs, extended context windows and new planning tools, Google said Gemini 3 can be applied to tasks such as analysing research papers, translating handwritten documents, generating visualisations, and evaluating sports performance. The company also said Search's AI Mode now includes generative UI features and interactive simulations.

Google rolls out Gemini 3 Deep Think to AI Ultra subscribers

Google is rolling out Gemini 3 Deep Think mode to Google AI Ultra subscribers in the Gemini app to enhance reasoning capabilities for complex math, science, and logic problems that challenge advanced models. This mode provides a substantial boost in handling intricate tasks. It excels on demanding benchmarks, achieving 41.0 percent on Humanity's Last Exam without tools and an unprecedented 45.1 percent on ARC-AGI-2 with code execution. These results position Gemini 3 Deep Think as a leader in the industry for rigorous evaluations. The technology relies on advanced parallel reasoning, which allows simultaneous exploration of multiple hypotheses. This approach builds directly on the capabilities of Gemini 2.5 Deep Think variants. Those earlier versions recently met gold-medal standards at the International Mathematical Olympiad and the International Collegiate Programming Contest World Finals, demonstrating high performance in competitive mathematical and programming challenges. Ultra subscribers gain immediate access to Gemini 3 Deep Think mode. They can activate it by selecting "Deep Think" in the prompt bar and choosing Gemini 3 Pro from the model dropdown within the Gemini app.

Gemini 3 Deep Think Model Is Now Available to These Users

Gemini 3 Deep Think is said to have scored 41 percent on HLE benchmark Google has now rolled out its Gemini 3 Deep Think artificial intelligence (AI) model to eligible users. The model was first mentioned by the company last month, when it released Gemini 3 Pro globally. Arriving just a fortnight later, the next generation of Google's reasoning model brings higher performance across logic, reasoning, mathematics, and science to break down niche and complex problems. The Mountain View-based tech giant said that the model will only be available to its most expensive subscription tier. Gemini 3 Deep Think Comes to Ultra Users In a blog post, the tech giant announced the release of the large language model (LLM), claiming it is designed to tackle complex problems that "challenge even the most advanced state-of-the-art models." It will only be available to those subscribed to the Google AI Ultra plan. Users have been curious about the Gemini 3 Deep Think ever since it was first announced last month. The main reason behind it was its claimed benchmark scores on two specific tests. The first is Humanity's Last Exam, where Google claims the model scored 41 percent without any tool use. For context, this test focuses on academic reasoning and is said to be the most difficult test for AI models. Deep Think's score is now the highest for the test, and by quite some margin. The second is a similarly challenging reasoning benchmark, ARC‑AGI‑2. It involves visual puzzles and requires code execution to solve them. Deep Think is claimed to achieve 45.1 percent, a result the company called "unprecedented." Notably, the test is also said to gauge how close a model is to artificial general intelligence (AGI). As per the tech giant, Gemini 3 Deep Think achieves this level of performance through a new technique called parallel reasoning. In this, the LLM ditches the traditional single chain of thought approach, and instead, the model explores multiple potential solution paths simultaneously. It then converges on the most coherent answer through multiple rounds of evaluating the response and refining it. Despite the release, the model is currently available to a very small group of users who are paying Rs. 24,500 a month for the Google AI Ultra subscription. They can access the model from Gemini's website or app by selecting "Deep Think" in the prompt bar and opting for Gemini 3 Pro in the model picker.

Google rolls out Gemini 3 Deep Think mode with its "strongest reasoning capabilities": CEO Sundar Pichai

Google has begun releasing Gemini 3 Deep Think mode to AI Ultra subscribers on the Gemini app. The feature is designed for demanding maths, science and logic tasks, supported by advanced parallel reasoning and strong benchmark scores. Users can enable the mode by selecting "Deep Think" and choosing Gemini 3 Pro. Google has started rolling out Gemini 3 Deep Think mode to Google AI Ultra subscribers using the Gemini app. In a post on X, CEO Sundar Pichai highlighted how the new mode can handle highly complex tasks, calling it a feature that "brings our strongest reasoning capabilities yet." He also shared Gemini's post on X, which showed an example of its ability to "simulate complex 3D architecture." According to a Google blog post, Deep Think mode is built for tough maths, science and logic problems that push the limits of current AI systems. The company said Gemini 3 Deep Think leads key benchmarks such as Humanity's Last Exam at 41.0% (without tools) and ARC-AGI-2 at 45.1% (with code execution). Google attributes this performance to advanced parallel reasoning, allowing the model to test several ideas at once. This builds on earlier Gemini 2.5 Deep Think versions that earned gold-medal results at both the International Mathematical Olympiad and the ICPC World Finals. In the X post by Gemini, it said, "With Deep Think mode, Gemini uses iterative rounds of reasoning, resulting in more refined, nuanced code that helps you visualize, prototype and experiment with noticeably richer detail." Ultra subscribers can start using Gemini 3 Deep Think mode now by choosing "Deep Think" in the prompt bar and selecting Gemini 3 Pro from the model menu.