Google DeepMind has announced significant updates to its Gemini 2.5 model series, introducing new capabilities for reasoning, audio interaction, and developer control. The update includes the experimental "Deep Think" mode for 2.5 Pro, native audio output features, and expanded computer use capabilities via Project Mariner.
- Gemini 2.5 Pro now leads WebDev Arena and LMArena leaderboards and serves as the top model for learning based on educational expert evaluations.
- The new "Deep Think" mode allows 2.5 Pro to consider multiple hypotheses, achieving high scores on USAMO and LiveCodeBench benchmarks.
- Gemini 2.5 Flash is now more efficient, using 20-30% fewer tokens while improving performance across reasoning, multimodality, and coding benchmarks.
- Native audio output enables conversational experiences with tone steering, emotion detection, and multi-speaker text-to-speech in over 24 languages.
- Project Mariner's computer use capabilities are being integrated into the Gemini API and Vertex AI for broader developer experimentation.
- Enhanced security safeguards protect against indirect prompt injections, making Gemini 2.5 the most secure model family to date.
- Developer tools now include thought summaries for transparency, extended thinking budgets for cost control, and native MCP support for open-source tool integration.
These updates aim to provide developers with more transparent, controllable, and capable tools for building complex applications while improving safety and efficiency across the Gemini ecosystem.