A major new Google DeepMind study has found that its Gemini 3 Pro model can genuinely shift people's beliefs and spending decisions, even without being explicitly taught manipulative techniques. The research, involving over 10,000 participants across the US, UK, and India, tested real money and commitments rather than just self-reported opinions.


Gemini 3 Pro shows manipulative behaviour in AI safety tests

Gemini 3 pro shows manipulative behaviour in tests 

<p>Google has announced a major upgrade to Gemini 3 Deep Think, its advanced reasoning mode designed for complex problem-solving. The updated version is now available only to users subscribed to Google AI Ultra, the company said.</p><p>Deep Think is built to handle difficult tasks in science, research and engineering, especially areas where data is unclear and there is no single correct answer. According to Google, the latest update was developed with help from scientists and researchers.</p><h2><strong>Gemini 3 Deep Think</strong></h2><p>Gemini 3 Deep Think is a specialised reasoning mode within Google’s Gemini AI system. Unlike regular AI chat features, Deep Think focuses on analysing complex problems that need deeper logic and step-by-step reasoning.</p><p>Google says the tool blends scientific knowledge with practical engineering use. This means it is not limited to theory but can also support real-world tasks, such as modelling systems through code or analysing large datasets.</p>

<blockquote class="twitter-tweet"><p lang="en" dir="ltr">Gemini 3 Deep Think is getting a significant upgrade. We’ve refined Deep Think in close partnership with scientists and researchers to tackle tough, real-world challenges.<br><br>And it’s pushing the frontier across the most challenging benchmarks, achieving an unprecedented 84.6% on… <a href="https://t.co/5503F4FKcD">pic.twitter.com/5503F4FKcD</a></p>&mdash; Sundar Pichai (@sundarpichai) <a href="https://twitter.com/sundarpichai/status/2022002445027873257?ref_src=twsrc%5Etfw">February 12, 2026</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>

<h2><strong>What’s new in the latest upgrade</strong></h2><p>With this upgrade, Google says Deep Think can now solve more advanced challenges across maths, science and programming.</p><p>According to Google’s internal and external benchmarks, the updated model:</p><p>Scored 48.4% on Humanity’s Last Exam, a test designed to measure the limits of advanced AI models</p><p>Achieved 84.6% on ARC-AGI-2, as verified by the ARC Prize Foundation</p><p>Reached an Elo rating of 3455 on Codeforces, a competitive programming platform</p><p>Delivered gold medal-level performance at the International Math Olympiad 2025 benchmark</p><p>Google said performance has also improved in chemistry, physics and other scientific fields.</p>

<blockquote class="twitter-tweet"><p lang="en" dir="ltr">Today, we’re releasing a significant upgrade to our specialized reasoning mode, Gemini 3 Deep Think.<br><br>Deep Think is built to drive practical applications, enabling researchers to interpret complex data and engineers to model physical systems through code.<br><br>With the updated Deep… <a href="https://t.co/jupp49pvmw">pic.twitter.com/jupp49pvmw</a></p>&mdash; Google Gemini (@GeminiApp) <a href="https://twitter.com/GeminiApp/status/2021985731577852282?ref_src=twsrc%5Etfw">February 12, 2026</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>

<h2><strong>How Deep Think can be used in real life</strong></h2><p>Google highlighted several practical uses for Deep Think. Researchers can use it to study complex data, while engineers can model physical systems by writing and testing code.</p><p>One example shared by Google shows how users can turn a simple hand-drawn sketch into a 3D-printable design. Deep Think analyses the drawing, creates a digital model and generates a file that can be used for 3D printing.</p><h2><strong>Google AI Ultra subscribers</strong></h2><p>The upgraded Gemini 3 Deep Think is currently available only through the Gemini app for Google AI Ultra subscribers. Google has not said when, or if, the feature will be expanded to other users.</p><p>The move shows Google’s focus on positioning Gemini as a high-end AI tool for research, engineering and advanced problem-solving, as competition in the AI space continues to grow.</p>

Google has upgraded Gemini 3 Deep Think, its advanced reasoning mode for science, maths and engineering. The tool shows top-tier benchmark scores and is now available only to Google AI Ultra subscribers via the Gemini app.

Google has announced a major upgrade to Gemini 3 Deep Think, its advanced reasoning mode designed for complex problem-solving. The updated version is now available only to users subscribed to Google AI Ultra, the company said. Deep Think is built to handle difficult tasks in science, research and engineering, especially areas where data is unclear and there is no single correct answer. According to Google, the latest update was developed with help from scientists and researchers. Gemini 3 Deep Think Gemini 3 Deep Think is a specialised reasoning mode within Google’s Gemini AI system. Unlike regular AI chat features, Deep Think focuses on analysing complex problems that need deeper logic and step-by-step reasoning. Google says the tool blends scientific knowledge with practical engineering use. This means it is not limited to theory but can also support real-world tasks, such as modelling systems through code or analysing large datasets.

<p>Google has upgraded Gemini 3 Deep Think, its advanced reasoning mode for science, maths and engineering. The tool shows top-tier benchmark scores and is now available only to Google AI Ultra subscribers via the Gemini app.</p>

Gemini 3 Deep Think gets major upgrade, now limited to Google AI Ultra users

gemini-3

Gemini 3, News on Gemini 3, stories on Gemini 3, articles on Gemini 3, Gemini 3 stories at Wion, Gemini 3 news, Gemini 3 latest news

gemini 3

Gemini 3 Pro shows manipulative behaviour in AI safety tests

Gemini 3 pro shows manipulative behaviour in tests

Gemini 3 Deep Think gets major upgrade, now limited to Google AI Ultra users