.svg){kind=spacer}
Details
- Google DeepMind introduced DeepThink, a new reasoning mode for the Gemini 2.5 Pro AI model, at Google I/O 2025. It uses parallel thinking to evaluate multiple hypotheses at once, improving performance in math, coding, and multimodal tasks.
- The model scored highly on benchmarks: 2025 USAMO (math Olympiad), LiveCodeBench (coding), and MMMU (multimodal reasoning) with an 84% accuracy. CEO Demis Hassabis cited these results as a major step in AI reasoning.
- DeepThink is currently available to selected testers via the Gemini API, with wider access pending careful safety reviews. The rollout is being monitored with support from external experts.
- This upgrade builds on the March 2025 release of Gemini 2.5 Pro and avoids the need for costly methods like test-time ensembling. It extends earlier strategies such as chain-of-thought prompting and reinforcement learning.
- Alongside DeepThink, Google also previewed Gemini 2.5 Flash for fast tasks and Project Mariner agents that automate web tasks—both entering limited previews this summer.
Impact
DeepThink marks a significant move forward in AI reasoning by enabling more flexible and robust hypothesis evaluation. With applications in STEM and software fields, it positions Google DeepMind as a key player advancing AI utility, while its phased release reflects a measured approach to responsible innovation.
