- The performance continues to leap
- Response style optimized
- # Add function: Thinking Budgets
- The way of access
Google released an upgrade preview version of Gemini 2.5 Pro, which will be launched as ** enterprise level generic version (GA)** in the coming weeks ** “More intelligent, faster, more stable, more creative.”** This is the position of Google for its new generation of big models, Gemini 2.5 Pro. Not only did it top out in multiple AI rankings, it also opened up for the first time the “thinking budget” function for developers, making AI no longer a “black box” but a manageable, customized smart tool.
The performance continues to leap
The latest Gemini 2.5 Pro model has made significant progress in several authoritative evaluations:
-
Elo scores have risen by 24 points to reach 1470 on the language understanding competency assessment platform LMalena and remain at the top of the list;
-
In the Web programming capacity test WebDevArena, the score rose by 35 points to 1443, with the same number first;
-
Continuing to lead the way in coding capacity, particularly in the difficult multi-linguistic programming test Aider Polyglot;
-
At the same time, the highest levels are shown in rigorous comprehensive assessments, such as GPQA and Humanity’s Last Exam (HLE), which mainly examine the capabilities of the models in the fields of mathematics, science, common sense and reasoning**.
-
**LMARENA I)
-
** Web programming and front-end development capability** (WebDevArena I)
-
** Complex, multilingual code generation** (Aider Polyglot)
-
The ability to solve problems in academic/rative categories (GPQA + HLE)
Response style optimized
In particular, Google optimizes the model ‘ s expression style and format control for user feedback:
-
** More structured structure**: The answers are clearer and suitable for generating reports, instructions, curricula.
-
** Stronger creative expression**: support for diverse needs such as literary style, content rewriting, advertising, etc.
-
** Response format is more professional**: Lists, code blocks, references can be generated with structures, more suitable for engineering scenarios.
This means that Gemini is not just a “talker”, but ** knows what you’re going to do** and gives ** a more usable result**.
# Add function: Thinking Budgets
This new feature is the current Gemini unique control mechanism, which means:
-
You can define how much to think about each call to AI (calculating resource consumption);
-
Balance ** Quality of generation, response time and cost**;
-
Very suitable for tool-type products, cost-sensitive applications.
The way of access
The developers now have access to Gemini 2.5 Pro in two ways:
Google AI Studio
Vertex AI (Google Cloud)
Gemini 2.5 Pro integrated Gemini application (Android / iOS / Web)
