Google’s Gold-Medal AI Brain: Deep Think Goes Public—But Only for a Price

In a bold move blending elite AI achievement with exclusive access, Google has launched a public-facing version of its award-winning AI model, Deep Think, which recently stunned the world with its performance at the 2025 International Math Olympiad (IMO). This high-performance model—based on the same architecture that earned top scores against some of the most challenging problems in mathematics—is now available to premium Gemini Ultra subscribers. But there’s a catch: the cutting-edge brainpower comes at a cost of \$250 per year, or \$125 for an introductory three-month period.

The release of Deep Think reflects a growing trend where breakthrough AI capabilities once confined to research labs are being monetized and gradually rolled out to the public. More than a consumer tool, this model represents a step toward integrating academic-level problem solving into everyday applications—from science and math to design and code.

the Original

Google has officially made a version of its Math Olympiad-winning AI, Deep Think, available to the public via the Gemini app—though only for its Ultra-tier subscribers. This premium tier costs \$250 annually, or \$125 for a three-month trial. While the public version doesn’t quite match the gold-standard capabilities of the original model used in the IMO, it reportedly performs at a bronze level based on 2025 IMO benchmarks, making it highly capable for most advanced users.

Deep Think leverages a technique known as “parallel thinking,” allowing it to generate and process multiple ideas at once, integrating them into a unified solution. This capacity is supported by extended “thinking time” during inference, enabling deeper problem exploration, and enhanced reinforcement learning strategies that help the model improve its reasoning skills over time.

In real-world applications, Deep Think excels at iterative design, complex scientific and mathematical exploration, and even advanced coding tasks. It has also demonstrated strong performance in comprehensive, multi-disciplinary tests like Humanity’s Last Exam, which spans over 100 subjects.

When compared to Gemini 2.5 Pro, the Deep Think variant offers better content safety and a more objective tone—although it’s been noted to reject more benign requests, possibly due to its cautious safety protocols.

To use the model, subscribers must activate Deep Think through the Gemini app’s model selector. Google also plans to expand access to developers and academic testers via the Gemini API in the near future. Meanwhile, the top-performing version used in the IMO will be reserved for a select group of mathematicians and researchers to explore its academic potential further.

What Undercode Say:

Google’s release of Deep Think is more than just a flashy tech announcement—it’s a powerful statement about the direction of AI accessibility, research commercialization, and the gamification of intelligence. By tying a public rollout to an elite competition like the International Math Olympiad, Google has positioned itself not just as a tech innovator, but as an educational and intellectual partner—albeit one with a steep entry fee.

From a market perspective, this is a strategic play that taps into rising demand for intelligent assistants that go beyond text prediction. Users are hungry for AI that can genuinely think—not just autocomplete. With Deep Think, Google delivers on that promise by offering a version of AI that mirrors human-like problem-solving behaviors, such as brainstorming, evaluating options, and synthesizing knowledge across disciplines.

But the exclusivity of this release raises questions. At \$250/year, Deep Think clearly targets professionals, academics, and power users—leaving casual or lower-income users outside the gate. This creates a class divide in access to high-functioning AI, where those who can afford to pay are given a measurable cognitive advantage in coding, research, and analytical tasks.

Interestingly, Google’s choice to emphasize “parallel thinking” echoes human cognitive strategies, such as lateral thinking or systems-based design. This could signal a broader move in AI architecture away from linear pattern matching and toward genuine cognitive mimicry.

Another notable aspect is the safety tuning. While Deep Think is more cautious and objective, its higher rate of request refusals may frustrate users expecting flexibility. Still, for academic and high-stakes environments, this conservatism is arguably a feature, not a bug—helping prevent hallucinations or biased outputs.

On the technical side, the mention of Humanity’s Last Exam suggests Google is preparing for an era where AI is benchmarked not just on narrow tasks but on general, cross-domain cognition. This echoes AGI ambitions, where models must perform not just in math or language but across the human knowledge spectrum.

Ultimately, Deep Think is both a marvel and a mirror. It shows us what AI can do—but also reflects our growing dependence on these tools for high-level reasoning, creativity, and exploration. The risk? As with any powerful tool, those who control access shape the direction of progress. And right now, that access is expensive.

🔍 Fact Checker Results:

✅ Google’s Deep Think model did in fact score at gold-medal levels on IMO-style math benchmarks.
✅ The version released to the public performs at a bronze-level on 2025 IMO simulations.
❌ Deep Think is not yet available via API for general developers—only selected testers will receive access soon.

📊 Prediction:

Within the next 18 months, Deep Think or its successors will likely be embedded into educational platforms, engineering suites, and scientific software tools—making advanced reasoning AI a standard co-pilot for professionals. If pricing models adapt, we may also see AI-powered tutoring and problem-solving become a global resource, potentially disrupting both online education and STEM research workflows.