Google upgrades Gemini Deep Think with stronger maths, science and coding performance

Google Gemini

• Google upgrades Gemini 3 Deep Think for advanced problem-solving.
• Deep Think now excels in math, science, and engineering tasks.
• Early access to Deep Think via Gemini API for select users.

Did our AI summary help?

Google has announced a major upgrade to Gemini 3 Deep Think, its specialised reasoning mode built to handle complex problems across science, research, and engineering. The company says the new version pushes the frontier of AI intelligence, particularly in domains where problems have no clear rules, incomplete data, or no single correct answer.

Google said the updated Deep Think was developed in close collaboration with scientists and researchers, with a focus on real-world research challenges rather than abstract theory. The aim, according to the company, is to blend deep scientific understanding with practical engineering usefulness, allowing the system to move from theoretical reasoning into applied problem-solving.

"Gemini 3 Deep Think is getting a significant upgrade. We’ve refined Deep Think in close partnership with scientists and researchers to tackle tough, real-world challenges," said Sundar Pichai, CEO, Google, in a post on X.

The upgraded Deep Think is now available in the Gemini app for Google AI Ultra subscribers. For the first time, Google is also opening access through the Gemini API, initially to a select group of researchers, engineers, and enterprises as part of an early access programme.

Stronger reasoning and benchmark results

Google claims the new Deep Think sets new performance highs on several demanding benchmarks used to evaluate frontier AI systems. The model achieved a score of 48.4 percent without tools on Humanity’s Last Exam, a benchmark designed to test the upper limits of modern reasoning models. It also reached 84.6 percent on ARC-AGI-2, a result verified by the ARC Prize Foundation.

In competitive programming, Deep Think reportedly attained an Elo rating of 3455 on Codeforces, while also reaching gold medal-level performance on the International Math Olympiad 2025. Google said these results build on earlier demonstrations where specialised versions of Deep Think showed success at math and programming competitions.

Expanding into science and engineering

Beyond maths and coding, Google says Gemini 3 Deep Think now performs strongly across scientific disciplines such as chemistry and physics. The model achieved gold medal-level results on the written portions of the 2025 International Physics Olympiad and Chemistry Olympiad, and scored 50.5 percent on the CMT-Benchmark, which focuses on advanced theoretical physics.

Google also positioned Deep Think as a practical tool for engineers and researchers. The company said the system can help interpret complex datasets and model physical systems through code, with a particular emphasis on deploying these capabilities through the Gemini API so they can be integrated into real workflows.

Access and availability

Google AI Ultra subscribers can access the upgraded Deep Think mode starting today in the Gemini app. Researchers, engineers, and enterprises interested in using Deep Think via the Gemini API can apply for early access through Google’s programme.