Article illustration 1

Two weeks after Google and OpenAI showcased their AI models' medal-winning performances at the International Math Olympiad (IMO), Google is bringing a version of its champion model to consumers. Dubbed Deep Think, this new capability is now available exclusively to Gemini Ultra subscribers—those paying $250 annually or $125 for the first three months.

While not the full gold-medal-winning model (reserved for academic researchers), Deep Think represents a significant leap. Internal benchmarks show it achieves bronze-level performance on the 2025 IMO test suite. Crucially, Google engineered it for practicality: "It is faster for everyday tasks" than its competition-grade sibling, according to the announcement.

The Engine Behind the Performance

Deep Think's prowess stems from three technical innovations:
1. Parallel Thinking: Allows simultaneous generation and processing of multiple solution paths, dynamically combining ideas for optimal answers.
2. Extended Inference Time: Deliberately increased "thinking time" enables deeper exploration of complex problems before finalizing responses.
3. Advanced Reinforcement Learning: Continuously improves problem-solving strategies through sophisticated training techniques.

Article illustration 2

Caption: Google's diagram illustrates Deep Think's iterative approach to complex tasks like design and development.

Benchmark results shared by Google highlight Deep Think's strengths:
- Humanity's Last Exam: Excelled in this multimodal benchmark spanning 100+ subjects (math, science, humanities).
- Coding & Technical Design: Demonstrated superior iterative development capabilities.
- Safety & Objectivity: Showed improved content safety and tone objectivity compared to Gemini 2.5 Pro, though with a noted increase in benign request denials.

Access and Implications

Ultra subscribers can access Deep Think via a toggle in the Gemini app, albeit with a daily prompt limit. Google plans API access for trusted testers soon, while the full gold-medal model remains restricted to mathematicians and academics for research advancement.

This tiered release strategy highlights a growing trend: specialized AI models for specialized users and tasks. Deep Think isn't just a smarter chatbot; it's a tool engineered for developers, scientists, and engineers tackling thorny technical challenges. Its success—and limitations—will shape expectations for the next generation of premium AI assistants.

Source: Based on reporting by Sabrina Ortiz for ZDNET.