Google rolls out powerful creative problem-solving AI model Deep Think to the Gemini app

Google DeepMind, Alphabet Inc.’s artificial intelligence research arm, today announced the rollout of Gemini 2.5 Deep Think, a new creative problem-solving AI model.

The company stated the model is capable of addressing complex problems by considering multiple ideas simultaneously and choosing the best answer.

The new tool will be available on the Gemini app for Google AI Ultra subscribers for $250 per month.

Google said the model extends the Gemini model’s parallel “thinking time.” Similar to how people explore complex problems by examining different angles, Deep Think works through problems by generating and simultaneously considering multiple variations. It can also combine different ideas over time before arriving at a conclusion.

“By extending the inference time, or ‘thinking time,’ we give Gemini more time to explore different hypotheses and arrive at creative solutions to complex problems,” the Deep Think team said.

The team also developed new reinforcement learning techniques, a type of training model for AI, to encourage the model to carefully consider problems. This enables Deep Think to be a more intuitive problem-solver, it added.

Deep Think can be used for iterative development and design, such as for creative projects or complex development tasks. The team noted that it also excels at research and mathematical study, making it an excellent tool for science, and is very good at tough coding problems where problem formulation can be a sticky subject.

“We’ve been impressed by Deep Think’s performance on tasks that require building something complex, piece by piece,” the team said. “For example, we’ve observed Deep Think can improve both the aesthetics and functionality of web development tasks.”

The new model’s release comes after an earlier version achieved the gold-medal standard at this year’s International Mathematical Olympiad. Though that version took hours to reason through complex math problems, today’s release is designed to be more usable for everyday creative thinking while providing Bronze-level performance.

The team said Deep Think’s performance stands out against other models in challenging benchmarks that measure coding, science knowledge, and reasoning capabilities. For example, without tool use, Gemini 2.5 Deep Think excelled in LiveCodeBench V6, which measures code performance, and Humanity’s Last Exam, a benchmark that challenges expertise across a broad variety of knowledge domains such as science and math.

Gemini app users with access to Deep Think can toggle it on within the prompt bar when selecting 2.5 Pro in the model dropdown. It automatically works with tools such as code execution and Google Search and can produce much longer responses.

The company said it’s working to release Deep Think, with and without tools, to trusted testers via the Gemini application programming interface in the coming weeks.

Images: SiliconANGLE/Microsoft Designer, Google