Google’s highest tier Gemini model updates with science and research focus
Google’s Gemini 3 Deep Think update is here, specializing in science, research, and engineering. Now on the app and API for Ultra users and select enterprises.
Google has announced its latest update to Gemini 3 Deep Think, which has a focus on science, research, and engineering.
The reasoning mode is tailored toward users that have highly specialized use cases, which require clear guardrails, accurate data, and practical applications. The new Deep Think update is now available in the Gemini app for Google AI Ultra subscribers. Google is also making the AI mode available through the Gemini API to select researchers, engineers and enterprises for the first time.
Google shared that early testers of Gemini 3 Deep Think include Lisa Carbone, a Rutgers University mathematician who used Gemini 3 Deep think to find a logical law in her technical mathematics paper that had previously passed through human peer review without being flagged.
Other uses for the updated model allow users to create 3D-printed models based on their sketches. It generates a file that can be sent to a 3D printing machine for creation.
Google has showcased the prowess of Gemini 3 Deep Think in various academic benchmarks, which have exceeded the performance of previous versions of the technology. The latest version has set a new standard of 48.4%, without tools on Humanity’s Last Exam, a test fashioned to assess modern frontier models.
The model also scored a record 84.6% on the ARC-AGI-2 benchmark for reasoning tasks, which was verified by the ARC Prize Foundation. It achieved an Elo of 3455 on Codeforces, a platform that hosts competitive programming challenges. The model gained gold-medal results in math, chemistry, and physics at the 2025 International Olympiad for each subject.
The updates come after Google originally introduced Gemini 3 Deep Think in November 2025. The advanced AI model finds competition in similar brands, including OpenAI’s ChatGPT and Anthropic’s Claude, which both have higher tiered versions with more elaborate reasoning models.