Two weeks after Google and OpenAI showcased their models' medal-winning prowess at the International Math Olympiad (IMO), Google is democratizing a piece of that magic—for a steep price. The company has launched Deep Think, a specialized version of its Gemini 2.5 AI, now accessible to Gemini Ultra subscribers at $250 per year (or $125 for the first three months). While not the full gold-medal model that stunned mathematicians, this iteration delivers accelerated performance for real-world tasks like coding and complex reasoning, positioning it as a powerhouse for developers and researchers.

Article illustration 1

Google's Deep Think in action, demonstrating iterative design capabilities. (Image: ZDNet)

The Engine Behind the Excellence

Deep Think builds on the architecture that earned Google a gold-standard finish at the IMO, but optimizes it for practical use. Internal benchmarks show it achieves bronze-level performance on the 2025 IMO scale—a feat enabled by three key innovations:
- Parallel Thinking: The model generates and processes multiple ideas simultaneously, dynamically combining concepts to refine solutions. This mirrors human brainstorming but at machine speed.
- Extended Inference Time: Unlike standard models that rush responses, Deep Think "thinks longer," exploring more pathways before settling on high-confidence answers.
- Reinforcement Learning: Continuous self-improvement cycles allow it to evolve into a sharper problem-solver over time, particularly in math, science, and iterative design.

Google claims this trio of advancements propels Deep Think past its predecessor, Gemini 2.5 Pro, in benchmarks like Humanity's Last Exam—a multimodal test spanning 100+ subjects. Early results indicate notable gains in coding efficiency and scientific discovery, though with a caveat: stricter safety filters may reject benign queries more often.

Access and Implications

For now, Deep Think is a premium perk. Ultra subscribers get limited daily prompts via the Gemini app (toggle "Deep Think" in the model selector), with API access for trusted testers rolling out soon. The full gold-medal model remains exclusive to academics, aimed at advancing mathematical research—a move Google hopes will feedback into public offerings.

This launch intensifies the AI subscription wars, following OpenAI's GPT-5 teasers. At $250/year, Deep Think targets professionals who value elite problem-solving over affordability. Developers gain a tool that could revolutionize debugging or algorithm design, yet the cost barrier risks excluding indie creators and smaller firms. As one Google insider noted: "It’s not just about solving equations—it’s about who gets to solve them first."

For the tech ecosystem, Deep Think underscores AI’s shift from general chatbots to specialized, high-stakes instruments. But with great power comes great pricing—sparking debates on whether innovation should be a luxury or a right.

Source: Based on reporting by Sabrina Ortiz for ZDNet. Original article: Google releases its award-winning Math Olympiad model.