Major update to Gemini 3 Deep Think: A new era in science and engineering

Google announced a broad update for Gemini 3 Deep Think, the advanced reasoning mode of the Gemini family. According to the company's message, the focus is clear: producing more reliable results on scientific and engineering problems that are ambiguous, multi-step and don't have a single correct answer.

This update isn't limited to raising the model's theoretical capacity. Google aims to speed up the transition "from research to application" by connecting Deep Think's performance on academic tests to workflows that are usable in real life.

The bar has been raised on hard problems

The new version of Deep Think stands out especially on benchmarks that require complex reasoning. According to the results shared, the model:

• Reaches 48.4 percent on Humanity's Last Exam (without tools),
• Scores 84.6 percent on ARC-AGI-2, verified by the ARC Prize Foundation,
• Climbs up to 3455 Elo on Codeforces.

This momentum in math and algorithms also carries over to the natural sciences. Google says the model produced results close to gold-medal level on the written sections of the 2025 International Physics and Chemistry Olympiads. Its 50.5 percent score on CMT-Benchmark, which focuses on theoretical physics, is among the indicators supporting this claim.

Not just exam success, but engineering output

The most notable aspect of the update is that the model is becoming more visible in real-world work. Deep Think offers more practical suggestions on tasks such as making sense of messy research data, modeling physical systems in code and managing multiple constraints at once.

Another scenario Google highlights is a "sketch to production" pipeline. A sketch prepared by the user can be analyzed by the model and turned into a three-dimensional model; after that, a file suitable for a 3D printer can be generated. This is an important step in turning a design at the idea stage into a physical prototype in a short time.

The access model is expanding

Starting today, Deep Think mode is available in the Gemini app for Google AI Ultra subscribers. In addition, Google is making this capability accessible through the Gemini API for the first time.

Under the early access program, researchers, engineering teams and enterprise users will be able to test the system in their own workflows. This step could pave the way for Deep Think to be used more actively not only on the consumer side but also in product development and R&D processes.

Assessment

The Gemini 3 Deep Think update claims to close the gap between "benchmark success" and "field use." Going forward, the real deciding factor will be how consistently these high scores translate into day-to-day engineering productivity, product quality and decision-making speed.

> Note: This content was prepared in an original form based on public announcements and shared technical results.