Google upgrades Gemini 3 Deep Think for science, research, and engineering

What changed

On Feb. 12, 2026, Google announced a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode for science, research, and engineering. Deep Think is now available in the Gemini app for Google AI Ultra subscribers, and Google says researchers, engineers, and enterprises can request early access to test it via the Gemini API for the first time. Google is clearly trying to move Deep Think from an internal or premium capability into a real product surface for research and R&D workflows.

Google says it trained and tuned the updated mode with scientists and researchers to tackle problems that do not have clear guardrails, clean datasets, or a single correct answer. The company highlights early examples from Rutgers University, where Deep Think identified a subtle logical flaw in a mathematics paper, and Duke University, where it helped design a crystal-growth recipe for larger thin films used in semiconductor materials research. Those examples show how Google wants the model to sit between abstract reasoning and practical engineering utility.

Benchmark leap and market signal

Google attached unusually ambitious numbers to the release. According to the company, Deep Think set a 48.4% score on Humanity’s Last Exam without tools, reached 84.6% on ARC-AGI-2, posted a 3455 Elo on Codeforces, and delivered gold-medal-level performance on the International Math Olympiad 2025. Google also says it reached gold-medal-level results on the written sections of the 2025 International Physics Olympiad and Chemistry Olympiad, plus 50.5% on CMT-Benchmark for advanced theoretical physics.

• Available now in the Gemini app for Google AI Ultra subscribers
• Early API access opened to researchers, engineers, and enterprises
• Google is positioning the model for both scientific research and practical engineering tasks

The significance goes beyond benchmarks. Google is using Deep Think to argue that frontier reasoning models can move from abstract math contests into practical research and product development. The launch examples span paper review, materials science, and physical component design, showing how Google wants Deep Think to sit between pure research tooling and commercial engineering software.

As with any vendor release, the headline results are Google’s own presentation of the model. Still, the Feb. 12 announcement stands out because it turns a specialized reasoning mode into an accessible product surface and an API program. That is a stronger commercialization signal than a research demo alone, especially for organizations looking at AI-assisted discovery and R&D workflows.