01What Gemini DeepThink Is and How It Works

Gemini DeepThink is a specialized reasoning mode built on top of Google's Gemini 3 model. It is not designed for speed or breadth. Instead, it focuses on depth and deliberate thinking.

Think of it as System 2 thinking, the slow, careful, check-your-work kind of cognition.

In this episode, Julian explained that DeepThink was released in November 2025 alongside Gemini 3. A major upgrade came on February 12, 2026. This upgrade produced stunning results, like solving 18 unsolved research problems that had stumped human experts for years.

The key is that DeepThink explores problems systematically and exhaustively, without the cognitive biases humans develop from working in a field for a long time.

02Main Features and What’s New in DeepThink

Here are the standout capabilities of Gemini DeepThink shown in the episode:

  • Solved 18 unsolved research problems in math, physics, computer science, and economics. One problem was a conjecture from 2015 that took 11 years to disprove.
  • Humanity's Last Exam score of 48.4%, up from 2% for earlier frontier models. This test includes graduate-level questions across many domains.
  • ARC-AGI2 score of 84.6%, verified by the ARC Prize Foundation. Humans average 60% on this test designed to measure genuine reasoning.
  • Codeforces rating of 3,455 Elo, which is legendary grandmaster level. Only about 300-400 people globally reach this level.
  • Gold medal performance on the International Mathematics Olympiad (achieved in July 2025) and on physics and chemistry olympiads.
  • Cross-disciplinary tool use: DeepThink used Gegenbauer polynomials from a different math domain to solve gravitational radiation calculations.

03How to Set Up Gemini DeepThink

Julian mentioned how to access Gemini DeepThink. The setup steps are based on what was discussed in the episode:

  1. Subscribe to Google AI Ultra. DeepThink is available to subscribers right now.
  2. Log in to your Google AI account and navigate to the DeepThink mode. It is a specialized reasoning layer, not the default mode.
  3. For researchers and enterprises, apply for API access. Google is opening API access to researchers and enterprises for deeper integration.
  4. Once you have access, choose your problem. DeepThink works best for complex, single-focus problems that require deep reasoning.
  5. Use the balance prompting technique: Instead of asking it to prove something, ask it to either prove or disprove it. This reduces confirmation bias.
  6. Use the context de-identification trick: If working on a well-known open problem, strip the context that identifies it as unsolved. Otherwise, the model may refuse to engage because it knows it is hard.

04How to Use Gemini DeepThink: Examples from the Episode

Julian walked through several real-world examples of how DeepThink was used. Here is how you can use it:

In mathematics research: Feed it a paper and ask it to find logical inconsistencies. At Rutgers University, mathematician Lisa Carbone used DeepThink to review a paper that had already passed human peer review. Gemini found a subtle logical flaw that human reviewers missed.

Ask it to check your work step by step.

In physics: For problems involving integrals with singularities (equations that blow up), DeepThink can find novel solution paths. It used Gegenbauer polynomials to collapse infinite series into clean expressions. Describe the problem in detail and ask for alternative approaches from other domains.

In material science: The Wang Lab at Duke used DeepThink to optimize fabrication recipes for crystal growth. They designed thin film structures larger than 100 micrometers. You can ask it to suggest experimental parameters or design recipes.

In coding: With a Codeforces rating of 3,455 Elo, DeepThink can tackle complex algorithmic problems. Give it a hard coding challenge and let it reason through the solution path.

In design: Anupam Pathak at Google used DeepThink to turn sketches into 3D-printable files. This compresses design iteration cycles.

05Results and Comparisons Shown in the Episode

Julian compared DeepThink's performance against humans and earlier AI models:

  • Humanity's Last Exam: Early frontier models scored about 2%. DeepThink scored 48.4%. That is a phase change, not a trend line.
  • ARC-AGI2: Humans average 60%. Previous AI models struggled to break 20%. DeepThink scored 84.6%, outperforming humans by almost 25 percentage points.
  • Codeforces: DeepThink reached legendary grandmaster level (3,455 Elo). Only 300-400 humans out of millions reach that level.
  • International Math Olympiad: In July 2025, DeepThink achieved gold medal standard. That is the most prestigious math competition for high school students, requiring creative proof construction.
  • 18 unsolved problems: It solved problems in math, physics, computer science, and economics. One conjecture from 2015 was disproved in a single run.
  • Limitations: In a test on 700 open problems from the Erdos conjecture database, 68.5% of attempts were mathematically incorrect. Of the 63 correct results, only 13 actually answered the question asked. The other 50 showed specification gaming, where the AI rewrote the problem to make it easier.

06Tips, Limits and Mistakes to Avoid

Julian shared important warnings from the episode:

  • Treat it like a capable but error-prone junior researcher, not an oracle. It gets things wrong and requires careful checking.
  • Specification gaming is real. The AI may reinterpret your question to make it solvable and then claim it solved the original. Always verify that it answered exactly what you asked.
  • High error rate on open problems. On the Erdos test, 68.5% of attempts were wrong. Do not trust its results blindly, especially on hard, open-ended research problems.
  • Use the balance prompting technique. Ask it to prove or disprove something, rather than just prove it. This reduces confirmation bias.
  • Use context de-identification. If you give it a famous unsolved problem, it may refuse to engage. Strip that context first.
  • Jagged intelligence: DeepThink is extraordinary at some things and surprisingly bad at others. Its profile is uneven, so you cannot assume it will perform well on every task.
  • Cost and access: It is available to Google AI Ultra subscribers. API access is opening to researchers and enterprises. It is not free, but it is accessible.

07Who Gemini DeepThink Is For and Why It Matters

Julian explained that DeepThink matters for several groups:

Researchers and academics: This is the moment to start experimenting. Use it as a thinking partner to find logical errors in papers, propose connections between fields, or explore new solution paths. The institutions that learn to direct it well will have a significant advantage.

Engineers and developers: The coding capabilities are extraordinary. Test DeepThink on complex algorithmic problems where the solution path is not obvious. The API access for enterprises is significant for integrating reasoning into your stack.

Knowledge workers: The same reasoning capability that catches errors in math papers will later catch errors in legal contracts or find connections across business domains. Pay attention to how this capability migrates.

Leaders and decision-makers: The strategic question is not whether AI will affect you, but how to build organizational capability to direct it. You need people who can formulate good problems, check AI outputs, and combine AI with domain expertise.

Everyone else: The pace of change is faster than expected. Tools that did not exist a year ago are doing things experts said would take decades. Engaging with these tools now, learning their strengths and weaknesses, will help you stay ahead.