Google has announced significant updates to its Gemini AI model series, introducing new capabilities and performance improvements. The company revealed Gemini 2.5 Pro with „Deep Think,“ an experimental enhanced reasoning mode, and an improved 2.5 Flash version that offers better performance while using fewer computational resources.
Gemini 2.5 Pro and Deep Think
According to Google, Gemini 2.5 Pro continues to lead in coding capabilities and now tops several benchmark leaderboards including WebDev Arena with an ELO score of 1415 and all categories of LMArena, which evaluates AI based on human preference.
The new experimental „Deep Think“ mode enables the model to consider multiple hypotheses before responding. Google reports impressive results on difficult benchmarks:
- Strong performance on the 2025 USA Mathematical Olympiad (USAMO)
- Leadership on LiveCodeBench for competition-level coding
- 84.0% score on MMMU, which tests multimodal reasoning
Currently, Deep Think is available only to trusted testers via the Gemini API while Google conducts additional safety evaluations.
Gemini 2.5 Flash improvements
The updated Gemini 2.5 Flash, designed for efficiency and speed, has been improved across benchmarks for reasoning, multimodality, coding, and long context understanding. Google claims the model is 20-30% more efficient, using fewer tokens in evaluations while maintaining high performance.
The new 2.5 Flash is now available in Google AI Studio for developers, Vertex AI for enterprise users, and the Gemini app for consumers, with general availability planned for early June.
New capabilities across the Gemini family
Both models are receiving new capabilities:
Native audio output: The Live API now offers audio-visual input and native audio output for more natural conversational experiences. Users can control tone, accent, and speaking style.
Text-to-speech: New text-to-speech features support multiple speakers in over 24 languages with the ability to capture nuances like whispers.
Project Mariner: Computer use capabilities are being integrated into the Gemini API and Vertex AI, allowing models to interact with computer systems.
Enhanced security: Google reports significant improvements in protection against security threats like indirect prompt injection attacks.
Developer experience improvements: New features include thought summaries for better understanding of model reasoning processes and thinking budgets that allow developers to control token usage.
New subscription tiers
Google is also introducing two subscription plans:
- Google AI Pro ($19.99/month): Replaces and expands on Gemini Advanced, offering higher rate limits and special features.
- Google AI Ultra ($249.99/month, US only): Provides access to the most powerful models, highest rate limits, and early access to experimental AI products, including Deep Think when it launches.
Sources: Google, Google, VentureBeat