Google has introduced the Deep Think reasoning model for Gemini 2.5 Pro, utilizing cutting-edge research to consider multiple hypotheses before responding to queries.
Deep Think by Gemini achieved an 'impressive score' in evaluations using questions from a math competition, but Google is taking time for safety evaluations and expert input before wider release.
A more efficient Gemini 2.5 Flash model has been launched, optimized for speed and benchmarks show improvements in reasoning, multimodality, code, and long context. It will be generally available in early June.
Google is extending the 'Thinking Budgets' concept from 2.5 Flash to the advanced 2.5 Pro model, aiming to balance tokens spent with accuracy and output speed. Additionally, Project Mariner by Google is being integrated into the Gemini API and Vertex AI for wider experimentation by developers.