- # The preferred model of the educational scene
- Deep Think: Experimental “Deep Thinking” mode
- # # high-level logic
- The current state
- Gemini 2.5 Flash: Light and efficient “AI tool motor”
- # Use the channel
- Original audio capacity: Dialogue is more natural and emotional
- # New feature bright spot
- Project Mariner: Let Gemini control your computer
- The key power
- I’ve got a lot of security upscaling
- To defend against new attacks
At the Google I/O 2025 Congress, Google announced Gemini 2.5 Pro and Flash functional enhancements and introduced a number of breakthrough capabilities, including:
-
High-level reasoning Deep Think
-
Native audio interactive capability
-
Multilingual voice generation
-
Project Mariner
-
Security and development experience optimization
♪ ♪ ooh ooh ooh ooh ooh ♪ ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh ooh
-
Advanced academic and practical capacity: Take the highest ELO rating 1415 (website development task) in WebDev Arena row;
-
Multiple dimensions are at the top of the human preference rating for LMARENA;
1 million token context window:
- Support complex multiple rounds of dialogue, lengthy document processing, video-by-college understanding.
# The preferred model of the educational scene
-
Integration of LearnLM (trained in collaboration with educational experts): (b) Be better at explaining knowledge and guiding learning;
-
Beyond competitions such as GPT-4 and Claude in teaching dialogue, human evaluation.
-
Become one of the most appropriate generic models for learning and teaching scenes at present.
Deep Think: Experimental “Deep Thinking” mode
# # high-level logic
-
Deep Think is a new experimental feature of Gemini 2.5 Pro, supporting the model “Multi-supposition parallel reasoning”, simulating multi-path thinking before answering;
-
The following outstanding tasks have been carried out: USAMO 2025 (American Mathematics Olympics): leading results;
-
LiveCodeBench (Code Capability Competition Task): ranked first;
-
MMMU (multimodular reasoning): Accuracy 84.0%.
The current state
-
To “trusted developers” only;
-
More rigorous security assessments are under way and gradual liberalization is expected in the future.
Gemini 2.5 Flash: Light and efficient “AI tool motor”
♪ ♪ Faster, cheaper, smarter ♪
-
Design objectives: ** Low delay + High throughput + Low cost **;
-
Full speed in reasoning, multimodel processing, long text tasks;
-
Token use reduction 20-30%, significantly reducing reasoning costs.
# Use the channel
- Developmenters and the public have been made available through Google AI Studio, Vertex AI and Gemini App.
Original audio capacity: Dialogue is more natural and emotional
# New feature bright spot
-
Native Audio Output: Support natural voice generation, control tone, emotions, speech style;
-
Fitting 24+ languages to support seamless transliteration of multiple languages;
Text-to-Speech (TTS) upgrade:
- Producing two-fold voice for dialogue, softly nuanced, emotional ups and downs;
Live API Extension:
-
Affective Dialogue: recognition of user sentiment and matching of feedback;
-
Proactive Audio: Auto-shield background noise, smart judgement whether to respond.
Project Mariner: Let Gemini control your computer
The key power
-
The ability to introduce “simulating human manipulation computers”: clicking, filling in forms, web interaction;
-
The following enterprises have been tested in cooperation: Automation Anywhere, UiPath, Browserbase, etc.;
More API test privileges will be opened to developers during the summer.
I’ve got a lot of security upscaling
To defend against new attacks
-
Gemini 2.5 significantly enhance protection against “Indirect Injection”;
-
Enhancement of the system ‘ s robustness in the use of tools through new testing mechanisms;
-
It’s currently Google’s safest version of the model.
Google has also added to Gemini API the original SDK support for the definition of the model context protocol (MCP) to allow easier integration with open source tools.