Google just upgraded its latest AI model, Gemini 2.5.
Flash is an AI model that allows you to give it a “thinking budget.”
Developers can now calibrate how much thinking Google’s Gemini model does for any task.
Google just rolled out an upgraded version of its latest AI model, with a new feature letting you “turn thinking on or off.”
On Thursday, the tech giant rolled out an early version of Gemini 2.5 Flash, an updated version of the 2.5 model it released in March.
That model — a so-called “thinking” model — was dubbed Google’s most intelligent one to date, given its ability to reason through ideas before responding.
However, Google is now ready to let you choose how much this new model thinks. And if you really want to, you can tell it to stop thinking completely.
In a blog post, Google Gemini’s director of product management, Tulsee Doshi, said that developers can “set thinking budgets to find the right tradeoff between quality, cost, and latency.”
The new feature aims to address the intense processing and computing requirements of a new wave of “reasoning” models that have spurred interest across the AI industry, including OpenAI’s o3, released on Wednesday.
Google’s new model aims to ensure that its reasoning model uses only as much processing power as necessary and applies it only when needed.
Doshi noted that not all tasks require the same reasoning. For example, the reasoning needed to answer “How many provinces does Canada have?” is different from asking AI to calculate the maximum bending stress on a cantilever beam of particular dimensions, she said.
We’ve just given our most powerful workhorse model a big upgrade to Gemini 2.5 Flash. You can try it now in preview on https://t.co/lLpF8ToTVJ – yet another Gemini data point on the cost-performance pareto frontier! https://t.co/UTUQJ8QWro pic.twitter.com/pTHBPIVydM
— Demis Hassabis (@demishassabis) April 17, 2025
To allocate different levels of reasoning abilities to user queries, Google will allow developers to set a “thinking budget” that Doshi said will offer “fine-grained control” over the number of tokens — units of data — a model generates while operating.
The move to introduce a “thinking budget” also follows a wider shift in the industry to become more “efficient” in the use of computing power.
This followed the release of a reasoning model in January from Chinese startup DeepSeek that claimed to use less computing power.
Read the original article on Business Insider
