Google DeepMind shipped a major upgrade to Gemini 3 Deep Think, its specialized reasoning model designed for complex scientific and engineering challenges.

The enhanced model targets researchers, scientists, and engineers working on problems with incomplete data or unclear solutions. Google developed the upgrade in partnership with academic researchers to handle real-world research scenarios.

Google AI Ultra subscribers can access the updated Deep Think through the Gemini app immediately. For the first time, the company is offering API access to select researchers, engineers, and enterprises through an early access program.

Real-world research applications

Early testers demonstrate the model's capabilities across diverse fields. Lisa Carbone, a mathematician at Rutgers University, used Deep Think to review a technical mathematics paper and identified a logical flaw that human peer reviewers had missed.

At Duke University, the Wang Lab employed Deep Think to optimize crystal growth methods for semiconductor materials. The model successfully designed fabrication recipes for thin films larger than 100 micrometers, meeting precise targets that previous methods struggled to achieve.

Anupam Pathak, an R&D lead in Google's Platforms and Devices division, tested Deep Think for accelerating physical component design workflows.

The upgraded model achieves new benchmarks across academic tests. It scored 48.4% on Humanity's Last Exam without tools, reached 84.6% on ARC-AGI-2 as verified by the ARC Prize Foundation, and attained an Elo rating of 3455 on Codeforces competitive programming challenges.

Google previously demonstrated Deep Think's capabilities by achieving gold-medal performance at mathematics and programming world championships. The company has also deployed specialized Deep Think agents for research-level mathematical exploration.

The model represents Google's push into scientific AI applications, competing with OpenAI's o1 reasoning models and Anthropic's Claude for complex problem-solving tasks.

Researchers and enterprises can express interest in API early access through Google's application form. The company has not disclosed pricing or broader availability timelines for the API version.