Google unveils Gemini 3 Deep Think upgrade for complex science and engineering tasks

Google has launched an upgraded Gemini 3 Deep Think mode aimed at complex science, research, and engineering tasks. It is available to Google AI Ultra subscribers in the Gemini app, while developers and enterprises can request API access through a Vertex AI early access program.

DeepMind says Deep Think leads several benchmarks against models like Claude Opus 4.6, GPT 5.2, and Gemini 3 Pro Preview, especially on logical reasoning and coding. On multimodal reasoning, the gap is much smaller, with MMMU Pro results nearly tied between Deep Think and Gemini 3 Pro Preview, suggesting the upgrade targets abstract reasoning more than visual understanding.

DeepMind also reports gold medal level performance at the 2025 Physics and Chemistry Olympiads, and points to examples of the model being used in scientific work.