Google is on a roll lately with new model releases. Following the successful launch of the Gemini 3 family of models, including Gemini 3 Pro and Gemini 3 Pro Image, Google today announced the launch of Gemini 3 Deep Think mode. This mode uses even more compute and new technologies compared to the regular Gemini 3 Pro model to deliver even better performance.
Google claims that this new Gemini 3 Deep Think mode performs significantly better in reasoning capabilities designed to solve complex math, science, and logic problems. On Humanity’s Last Exam, one of the industry’s toughest benchmarks for AI models, Gemini 3 Deep Think mode scored 41%, setting a state-of-the-art performance. On ARC-AGI-2, another rigorous benchmark, this new mode scored 45.1% with code execution. On the GPQA Diamond scientific knowledge benchmark, this mode scored 93.8%, another state-of-the-art performance.
Gemini 3 Deep Think was able to deliver such strong results because it uses advanced parallel reasoning to explore multiple hypotheses simultaneously. Recently, variants of this Deep Think model achieved a gold-medal standard at the International Mathematical Olympiad and at the International Collegiate Programming Contest World Finals. In the IMO, the models had to solve the problems in two 4.5-hour exam sessions, with no access to tools or the internet, and write natural language proofs.
Gemini 3 Deep Think mode is now available for all Google AI Ultra subscribers in the Gemini app. If you are a Google Ultra subscriber, you can check out the Gemini 3 Deep Think mode today by selecting “Deep Think” and the Gemini 3 Pro model in the prompt bar.
Back in July, OpenAI also claimed that its experimental reasoning LLM achieved gold medal-level performance. However, this model has not yet been made public. Since Google has released its IMO Gold Medal-standard model to the public, it would not be surprising if OpenAI also does the same in the near future.