Google Releases Major Update to Gemini 3 Deep Think, Enhancing Scientific Research Capabilities

Google has launched a significant update to its Gemini 3 Deep Think model, transforming it from a basic reasoning tool into a comprehensive scientific collaborator. Developed in collaboration with Google’s scientists and engineers, the updated model is designed to tackle complex research challenges where instructions are unclear, data may be contradictory, and errors carry high costs. Unlike previous versions that primarily provided answers, the new Gemini 3 Deep Think assists in identifying errors in scientific papers, simulating physical processes through coding, converting sketches into 3D printable models, and optimizing semiconductor manufacturing. Performance metrics highlight its advancements: a 48.4% score without tools on the Humanity’s Last Exam test, setting a new record; an 84.6% score on the ARC-AGI-2 benchmark verified by the ARC Prize Foundation; a 3455 Elo rating on Codeforces; and gold-level results in the 2025 International Math Olympiad as well as written physics and chemistry competitions.

Resources:

• blog.google
• blog.google