Google’s new Gemini 3 Deep Think update pushes the boundaries of AI reasoning

Google is taking another significant leap forward in the AI space with a major upgrade to Gemini 3 Deep Think. This specialized reasoning mode is designed to move beyond the conversational nature of standard LLMs, focusing instead on solving high-level challenges in science, research, and engineering. While standard Gemini models are built for speed and breadth, Deep Think is engineered for depth, utilizing advanced “System 2” thinking to tackle problems that lack clear solutions or contain messy, incomplete data.

This update is launching today for Google AI Ultra subscribers within the Gemini app. For the first time, Google is also opening up access via the Gemini API to a select group of researchers and enterprises through an early access program. This shift toward a more deliberate, reasoning-focused model marks a new chapter for the Gemini era, prioritizing accuracy and logical rigor over the near-instantaneous but sometimes superficial responses we’ve grown accustomed to.

Shattering benchmarks in reasoning and science

The performance metrics shared alongside this update are nothing short of staggering. Gemini 3 Deep Think has set a new standard on “Humanity’s Last Exam”—a benchmark specifically designed to test the absolute limits of frontier models—scoring 48.4% without the help of external tools. Perhaps more impressive is its 84.6% score on ARC-AGI-2, a reasoning benchmark verified by the ARC Prize Foundation that emphasizes a model’s ability to adapt to entirely new tasks rather than relying on memorized patterns.

In the realm of pure mathematics and competitive programming, the model continues to dominate. It currently holds a 3455 Elo on Codeforces and has reached gold-medal level performance on the 2025 International Math Olympiad. These aren’t just abstract numbers; they represent an AI that can identify subtle logical flaws in technical papers that have slipped past human peer reviewers. For researchers, this means having a partner capable of verifying complex proofs and exploring theoretical physics at a level previously reserved for human experts.

From abstract theory to practical engineering

Google is making it clear that Deep Think isn’t just for academic ivory towers. One of the most compelling “real-world” applications highlighted in this release is the model’s ability to bridge the gap between a simple sketch and a physical object. The updated Deep Think can analyze a hand-drawn design, model the complex geometry required to make it functional, and generate a 3D-printable file. This level of spatial reasoning and engineering utility opens doors for rapid prototyping that were previously closed to AI.

Beyond physical modeling, the model is proving its worth in chemistry and physics, achieving gold-medal results on the written portions of the 2025 International Physics and Chemistry Olympiads. Whether it’s interpreting complex data sets or modeling physical systems through code, Google is positioning Deep Think as a “force multiplier” for human intellect, handling the heavy lifting of verification so that engineers and scientists can focus on creative direction.

Accessing the power of AI Ultra

This top-level subscription is becoming the home for Google’s most advanced experimental features, including the latest video generation tools and now, this enhanced reasoning mode. If you are a Google AI Ultra subscriber, you can head over to the Gemini app today to begin testing the Deep Think mode for yourself. For the developers and researchers among us, the expansion to the Gemini API is a massive step forward for building more reliable, agentic workflows that can actually reason through problems rather than just predicting the next word in a sentence.