A Google Gemini model now features a “slider” to control its level of reasoning.
“We’ve been really pushing on ‘thinking,’” says Jack Rae, a principal research scientist at DeepMind. Such models, which are designed to solve problems logically and spend more time arriving at an answer, gained attention earlier this year with the introduction of the DeepSeek R1 model. They are appealing to AI companies because they can enhance existing models by teaching them to tackle problems pragmatically, thus saving the time and resources required to build a new model from scratch.
When an AI model dedicates more time (and energy) to a query, it results in higher operational costs. Leaderboards of reasoning models indicate that completing a single task can exceed $200. The rationale is that this additional time and investment can lead reasoning models to perform better on complex tasks, such as analyzing code or extracting information from numerous documents.
“The more you can iterate over certain hypotheses and thoughts,” says Google DeepMind chief technical officer Koray Kavukcuoglu, “the more it’s going to find the right thing.” However, this isn’t universally beneficial. “The model overthinks,” explains Tulsee Doshi, who leads the product team at Gemini. This refers specifically to Gemini Flash 2.5, which features a slider for developers to limit the model’s reasoning capacity. “For simple prompts, the model does think more than it needs to.”
When a model spends excessive time on a problem, only to arrive at a mediocre response, it becomes costly for developers and negatively impacts AI’s environmental footprint. Nathan Habib, an engineer at Hugging Face who has studied the rise of such reasoning models, notes that overthinking is widespread. In the effort to showcase more intelligent AI, companies are employing reasoning models like tools even when they aren’t necessarily suited for the task, according to Habib. Indeed, when OpenAI unveiled a new model in February, it indicated that it would be the company’s final non-reasoning model.
The performance improvement is “undeniable” for certain tasks, Habib states, but not for many other applications where AI is typically utilized. Even when reasoning is applied correctly, issues can arise. Habib presented an example of a leading reasoning model that was tasked with solving an organic chemistry problem. While it started well, midway through its reasoning process, the model’s responses began to unravel: It repeated “Wait, but …” countless times. Ultimately, it took significantly longer than a non-reasoning model would on the same task. Kate Olszewska, who evaluates Gemini models at DeepMind, mentions that Google’s models can also become trapped in loops.
Google’s new “reasoning” dial is one effort to address this challenge. Currently, it is designed not for the consumer version of Gemini but for developers creating applications. Developers can allocate a budget for computing power the model should utilize for specific problems, with the intention of reducing the reasoning capacity for tasks that don’t require extensive reasoning. Outputs from the model become approximately six times more expensive when reasoning is activated.
