The GEMINI 2.5 Flash of Google provides “thinking budgets” that reduce the costs of artificial intelligence by 600 % when they are rejected

Google Firing Gemini 2.5 flashA great promotion to the formation of artificial intelligence that gives companies and developers unprecedented control over the amount of “thinking” that artificial intelligence leads. The new model, which was released today in a preview through Google Ai Studio and Ai headIt represents a strategic effort to provide improved thinking capabilities while maintaining competitive prices in the increasingly crowded artificial intelligence market.

The model offers what you call Google.Thinking budget– A mechanism that allows developers to determine the amount of mathematical energy that must be dedicated to thinking through complex problems before generating a response. This approach aims to address the basic tension in the artificial intelligence market today: it is usually the most advanced thinking at the expense of cumin and higher prices.

“We know that the cost and cumin are a matter of a number of developers use, so we want to provide flexibility developers in adapting the amount of thinking the model, depending on their needs,” said TULSEE DOSHI, the director of the Gemini Products at Google DeepMind.

This flexibility reveals the practical Google approach to spread artificial intelligence because technology becomes increasingly included in business applications as the ability to predict costs is necessary. By allowing to run or stop the possibility of thinking, Google created what you call “the first full hybrid thinking model”.

Pay only for the strength of the minds you need: inside the new AI formation model from Google

The new pricing structure highlights the cost of thinking about artificial intelligence systems today. When using Gemini 2.5 flashThe developers pay $ 0.15 per million icons of inputs. Production costs vary greatly based on thinking settings: $ 0.60 per million symbols while stopping thinking, jumping to $ 3.50 per million symbols with enabling thinking.

This six times the teams reflects nearly six times in logical outputs, the arithmetic intensity of the “thinking” process, where the model evaluates potential multiple paths and considerations before generating a response.

“Customers pay for any thinking symbols and remove the model generated by the model,” Duchi told Venturebeat. “In Ai Studio UX, you can see these ideas before responding. In the application programming interface, we are not currently providing access to ideas, but the developer can know the number of symbols created.”

Thinking budget can be modified from 0 to 24,576 symbols, which works as a maximum instead of fixed customization. According to Google, the model intelligently determines the amount of this budget that must be used based on the complexity of the task, and to maintain resources when the distinctive thinking is necessary.

How GIMINI 2.5 Flash accumulates: Standard Results against Seistic Models of Artificial Intelligence

Google claims Gemini 2.5 flash Competitive performance through the main criteria shows the size of a smaller model of alternatives. on The last humanity examA strict test designed to evaluate thinking and knowledge, and a 2.5 -flash record 12.1 %, outperform anthropology Claude 3.7 Sonata (8.9 %) and Deepsek R1 (8.6 %), although Openai has not launch recently O4-Mini (14.3 %).

The model also published strong results on technical standards such as GPQA diamonds (78.3 %) and Aime mathematics exams (78.0 % in tests 2025 and 88.0 % in 2024 tests).

“Companies should choose 2.5 flash because it provides the best value and speed,” Duchi said. “It is particularly strong for competitors in mathematics, multimedia thinking, long context, and many other major standards.”

Industry analysts note that these criteria indicate that Google narrows the performance gap with competitors while maintaining the pricing feature – a strategy that may resonate with institutional customers who see artificial intelligence budgets.

Smartly for a fast: When does artificial intelligence need to think deeply?

The insertion of adjustable thinking is a major development in how Amnesty International has published. With traditional models, users do not have a great degree in the process of internal thinking of the model.

Google’s approach allows developers to improve different scenarios. For simple information such as language translation or basic information recovery, thinking can be disabled to maximize cost efficiency. For complex tasks that require multiple -step thinking, such as solving mathematical problems or careful analysis, thinking function can be enabled and adjusted.

The main innovation is the model’s ability to determine the amount of thinking based on the query. Google explains this with examples: A simple question such as “How many provinces can Canada have?” The minimum thinking requires, while the complex engineering question about the ray stress accounts will automatically share deeper thinking processes.

“Including the possibilities of thinking about our main Gemini models, along with improvements in all fields, led to high -quality answers,” Douche said. “These improvements are correct through academic standards – including Simpleqa, which measures realism.”

Google Week of Artificial Intelligence: Free Students’ arrival and video generation joined the launch of the Flash 2.5

release Gemini 2.5 flash It comes within a week of aggressive moves by Google in the area of ​​artificial intelligence. On Monday, the company was published Veo 2 The possibilities of generating video for Gemini Advanced subscribers, allowing users to create eight -second videos of text claims. Today, along with the Flash 2.5 announcement, Google revealed this All university students in the United States will receive free arrival at Gemini Advanced until the Spring of 2026 It is a step that analysts explained as an attempt to build loyalty among future knowledge workers.

These ads reflect a multi -sided Google Strategy to compete in a market dominated by Openai, which was reported more 800 million weekly users Compared to the estimated Gemini 250-275 million monthly usersAccording to third -party analyzes.

Flash 2.5, with its explicit focus on cost efficiency and performance customization, seems to be specially designed for institutional customers who need to manage the costs of spreading artificial intelligence carefully while continuing to reach advanced capabilities.

“We are very excited to start getting notes from developers about what they are building with Gemini Flash 2.5 and how they use thinking budgets,” Douche said.

Beyond the inspection: What companies can expect with the maturity of Gemini 2.5 Flash

Although this version is in the inspection, the model is already available to developers to start construction, although Google has not set a schedule for general availability. The company indicates that it will continue to improve the capabilities of dynamic thinking based on the developer’s notes during this preview phase.

For institutions AI for institutions, this version represents an opportunity to experience more accurate methods of spreading artificial intelligence, which may allocate more arithmetic resources for high -risk tasks while maintaining costs on routine applications.

The model is also available for consumers through Gemini applicationIt appears to be “2.5 flash (experimental)” in the model drop -down menu, to replace the previous 2.0 (experimental) thinking option. This publication, which the consumer faces, indicates that Google uses the application’s environmental system to collect wider notes on the structure of thinking.

Since artificial intelligence is increasingly included in commercial work, the Google approach with customized thinking reflects a mature market as the cost improvement and performance control becomes the same importance as raw capabilities – which indicates a new stage in marketing artificial intelligence technologies.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.