Google Unveils Gemini 2.5 With Deep Think, Audio Output
At Google I/O 2025, the tech giant unveiled major updates to its Gemini 2.5 AI models. A standout feature is Deep Think mode, designed to improve how the AI reasons through complex tasks. Built into the Gemini 2.5 Pro model, Deep Think allows the AI to weigh multiple possibilities before answering. This represents a shift from previous methods, where decisions followed a single path of logic.
According to Google, internal testing showed Deep Think scored 49.4% on the UAMO 2025 benchmark—one of the toughest mathematical reasoning tests. It also performed well on LiveCodeBench v6 and MMMU. Although still in testing, Deep Think is currently available to trusted developers through the Gemini API. For now, a full rollout date remains unconfirmed.
Gemini Flash Sees Efficiency and Performance Gains
In addition to Pro, the Gemini 2.5 Flash model also received key upgrades. Google claims the model now consumes 20–30% fewer tokens, improving cost and processing efficiency. More importantly, Flash has made strides in reasoning, multimodal understanding, code generation, and long-context processing.
This updated version is available in preview through Google AI Studio and for enterprises via Vertex AI. Public availability is expected next month. Thanks to these changes, developers can expect smoother, more powerful AI outputs with reduced resource use.
Expressive Audio and Thought Transparency
Another exciting feature introduced is Native Audio Output. This tool enables Gemini to generate human-like speech with controllable tone, accent, and style. Three core capabilities support this: Affective Dialogue, which responds to user emotion; Proactive Audio, which filters ambient noise; and Thinking, where Gemini explains complex responses verbally.
Furthermore, Google now includes thought summaries in the Gemini API and Vertex AI. These offer developers insight into how the AI arrived at an answer, showing key points and internal steps. This move boosts transparency and builds trust.
To help manage processing power, Google will soon roll out thinking budgets for Gemini Pro. These allow developers to set limits on how many tokens an AI can use when solving a problem. Also coming soon is the Project Mariner Computer Use function, which adds agentic reasoning to API tasks.
With these features, Google positions Gemini 2.5 as a smarter, more efficient, and more human AI, just in time for the next generation of intelligent applications.
