Based on the continuous expansion of the Gemini 3 model family, Google has officially released the new Gemini 3 Deep Think mode. This mode invests more computing power and introduces new technologies on Gemini 3 Pro, aiming to provide a higher level of performance on complex mathematics, science and logical reasoning tasks.
Currently, Deep Think is open to all Google AI Ultra subscribers and can be used directly in the Gemini app by selecting the Gemini 3 Pro model in the prompt bar and switching to "Deep Think" mode.
According to reports, Gemini 3 Deep Think has achieved "industry-leading" results in a number of difficult benchmark tests in the industry. It achieved a score of 41% on Humanity’s Last Exam, considered one of the most difficult AI reasoning tests currently, a score of 45.1% on the ARC-AGI-2 test using code execution, and a high score of 93.8% on the GPQA Diamond scientific knowledge test, all marked as “state-of-the-art.” Google said that this model relies on "advanced parallel reasoning" capabilities to explore multiple hypothesis paths at the same time, thereby greatly improving accuracy and robustness in solving complex problems.
In actual competition scenarios, Deep Think related models have also demonstrated capabilities close to top human players. Google mentioned that these Deep Think variants recently reached "gold standard" levels at the International Mathematical Olympiad (IMO) and the International Collegiate Programming Contest (ICPC) World Finals. In the IMO test setting, the model needs to independently solve problems and write natural language proofs in two 4.5-hour exams, completely offline and without access to tools or the Internet, which requires extremely high levels of reasoning depth and expression ability.
For end users, Gemini 3 Deep Think is positioned as the "high reasoning mode" in the Ultra subscription tier, mainly for complex task scenarios that require rigorous deduction, multi-step reasoning and high reliability. Google emphasizes that this model is stacked on the existing Gemini 3 Pro capabilities through additional computing power and new technologies, so it is more suitable for handling difficult mathematics, scientific research problems and complex logical analysis needs. Users only need to switch to Gemini 3 Pro in the model options and enable Deep Think within the Gemini app to experience this enhanced inference capability.
The article also mentioned that OpenAI claimed as early as July this year that one of its large experimental reasoning models had reached "gold-level" performance, but the model has not yet been publicly released. In contrast, Google will directly open the Deep Think model that reaches the IMO gold standard to Ultra subscribers this time, which is regarded as a "preemptive release" of competing products. It is also widely expected that OpenAI may follow up in the near future and push its high-inference model to the public.